What’s new in Tempr

Added

• Turn web search off. Settings > Tools & agents now has a Web search switch, on by default. Off, the agent can still open a page, but not search the web for one. Each search is billed to the model provider's key, or to your Perplexity key on models whose provider has no search of its own (about $1 per 1,000 searches).

Added

• Turn web search off. Settings > Tools & agents now has a Web search switch, on by default. Off, the agent can still open a page, but not search the web for one. Each search is billed to the model provider's key, or to your Perplexity key on models whose provider has no search of its own (about $1 per 1,000 searches).

Added

• Turn web search off. Settings > Tools & agents now has a Web search switch, on by default. Off, the agent can still open a page, but not search the web for one. Each search is billed to the model provider's key, or to your Perplexity key on models whose provider has no search of its own (about $1 per 1,000 searches).

Security

• Safer connections to MCP servers. Tempr's MCP library is updated to 1.32.2, with fixes for its published advisories, among them sign-in details sent to a server the MCP server named and request headers sent on after a redirect to another site. The libraries it uses are updated too. MCP servers work as before.

Changed

• Clearer wording. Error and sign-in messages, slash-command help, What's new and the Marketplace page read more plainly, and so do the instructions the agent works from.

Changed

• Clearer wording. Error and sign-in messages, slash-command help and the release notes read more plainly, and so do the instructions the agent works from.

Changed

• Clearer wording. Error and approval messages, What's new and the Marketplace page read more plainly, and so do the instructions the agent works from.

Changed

• An empty model list says why. When you're signed in but haven't added a provider key yet, the model list says "No models yet: add a provider key" instead of opening to nothing. Its menu offers Add a provider key, which opens your provider keys in the Tempr Portal, and Check again, which fetches the list again once you've added one.
• The walkthrough says where provider keys go. Get Started with Tempr now links to Provider keys in the Tempr Portal, and says a single OpenRouter key works for Chat (the Tempr Gateway doesn't take OpenRouter keys).

Changed

• An empty model list says why. When you're signed in but haven't added a provider key yet, the model list says "No models yet: add a provider key" instead of opening to nothing. Its menu offers Add a provider key, which opens your provider keys in the Tempr Portal, and Check again, which fetches the list again once you've added one.

Fixed

• Inline completion comes back on its own. When it pauses after repeated failures, it now tries one request every 5 minutes and picks up again as soon as one works, so a short drop in your connection no longer turns it off for the rest of the session.
• Try again after a failed turn. A turn that failed now offers Try again, which runs it again without sending your message twice, and the error's Regenerate button does the same. A model list that failed to load offers Try again too, instead of asking you to reopen the view.
• Deleted chats stay deleted. A synced chat deleted while Tempr's server couldn't be reached came back on the next start. The delete is now kept and sent again until the server confirms it.
• Chat sync, custom agents and the MCP catalog stay signed in. They now renew your sign-in when it runs out, as the chat does, and give up after 30 seconds instead of waiting on a server that never answers. Signing in with a license key also gives up after 30 seconds and says so, and your custom agents are fetched again at start if the first try fails for a moment.
• A turn is never started twice. Starting a chat is now sent again only when the connection was refused or Tempr was busy, not after a server error that could follow a turn that had already started.
• Notices from Tempr are plain text. A notice can no longer carry links that run commands or open files; its only link is the Open button for its web page. A damaged notices file is ignored instead of breaking inline completion, and a pause Tempr sets for your version holds from the moment VS Code starts.

Changed

• Tempr can tell you about a problem in your version. When a version has a known problem, VS Code shows a short message with a link to the fix, at most once a day. If a feature misbehaves in one version, Tempr can pause it there (inline completion, uploads of your chats, automatic error reports) until an update fixes it, and tells you once.

Fixed

• You stay signed in. When your sign-in runs out, Tempr renews it quietly and sends the request again; if the server has ended it, Tempr signs in again with your saved license key. Only when both are refused does it ask you to re-enter your key. A moment without a connection no longer signs you out.
• A dropped connection is tried again. Requests that fail for a moment are tried again with growing waits, honoring the server's Retry-After, and a chat whose connection drops reconnects instead of ending. Starting a chat is only sent again when it can't have reached Tempr, so a turn is never started twice.
• Stop works right away. Pressing Stop before a turn has started now cancels it as soon as it does, and Stop ends a reconnect wait at once.
• MCP servers that fail to start are tried again, then named in one notice with Try again. Your organization's MCP tools keep working past 15 minutes, and an MCP error no longer leaves its process running.
• Inline completion pauses after repeated failures. After three failures in a row it pauses for the session, with a notice and Try again.

Changed

• Tempr can tell you about a problem in your version. When a version has a known problem, the chat shows a short notice with a link to the fix, at most once a day. If a feature misbehaves in one version, Tempr can pause it there (inline completion, uploads of your chats, automatic error reports) until an update fixes it, and tells you once.

Fixed

• You stay signed in. Tempr renews your sign-in on its own, also while the chat window is closed, so inline completion and inline chat no longer stop after 15 minutes with "Your license session has expired." A moment without a connection no longer signs you out or uses up an activation.
• A dropped connection is tried again. A request that fails for a moment, because the server is busy or the connection dropped, is tried again before you see an error, and a chat whose connection drops picks up where it left off. Starting a chat is only sent again when the first try never reached Tempr, so a turn is never started twice.
• When the chat view stops working, it says so. If the chat view's browser process ends, the chat area says it stopped working and offers Try again, which starts it afresh. The settings screen gets the same retries as the chat.
• Inline completion pauses after repeated failures. After three failures in a row it pauses for the session, with a notice and Try again, instead of failing on every pause in typing.
• Fewer ways for Tempr to disturb Visual Studio. Clicks and Visual Studio automation calls that could throw are now caught, and a failed error report no longer files another one.

Fixed

• Visual Studio no longer closes when Tempr's chat view can't start. If WebView2, which shows the chat, failed to start, as can happen while another Visual Studio window is closing, Visual Studio itself shut down. Tempr now tries again twice, and if the chat view still can't start, the chat area says so and offers Try again. Your chats are kept and come back when it starts.
• Error reports are only about Tempr, and say which version sent them. Automatic error reports also picked up crashes in other extensions, and always gave the version as 1.0.0.0. They now cover only Tempr's own errors and carry the version you're running.

Fixed

• Signing out ends your session on the server. "Tempr: Sign Out" used to remove the sign-in from this machine only, so a copy of it kept working until it expired. It now ends the session on Tempr's server too. If the server can't be reached, it still signs this machine out and says so.

Fixed

• Signing out ends your session on the server. Sign out in Tempr's settings used to remove the sign-in from this machine only, so a copy of it kept working until it expired. It now ends the session on Tempr's server too. If the server can't be reached, it still signs this machine out and says so.

Fixed

• Deleting your license key ends your session on the server. Deleting the saved license key used to remove the sign-in from this machine only, so a copy of it kept working until it expired. It now ends the session on Tempr's server too. If the server can't be reached, it still signs this machine out and says so.

Changed

• Ask for a tool your organization's Tool Packs leave out. When a tool on one of your organization's MCP servers isn't in your Tool Packs, the tool call's "Not allowed by your organization" now comes with a Request access button, which opens a Portal page in your browser where you ask an owner or admin for it. Once they approve, your next turn can use it. The agent is told only that the tool can't be used, as before.
• Access changes reach an open chat. A tool an owner or admin approves for you, or a change to your Tool Packs, now shows up on the next turn of a chat that's already open, without a reload.

Changed

• Ask for a tool your organization's Tool Packs leave out. When a tool on one of your organization's MCP servers isn't in your Tool Packs, its card now shows a Request access button next to "Not allowed by your organization". It opens a Portal page in your browser where you ask an owner or admin for the tool; once they approve, your next turn can use it. The agent is told only that the tool can't be used, as before.
• Approved tools reach an open chat. A tool an owner or admin approves for you, or a change to your Tool Packs, now shows up on the next turn of a chat that's already open, without restarting the IDE.

Changed

• Ask for a tool your organization's Tool Packs leave out. When a tool on one of your organization's MCP servers isn't in your Tool Packs, the tool now shows a Request access button under "Your organization's Tool Packs don't include this tool." It opens a Portal page in your browser where you ask an owner or admin for the tool, and once they approve, your next turn can use it. The agent is told only that the tool isn't available, as before. An approved tool now shows up on the next turn of a chat that's already open.

Added

• Your organization's rules and skills. When your organization publishes rules and skills for its agents, the agent follows its rules on every turn, next to your project's own instructions, and reads one of its skills when a task matches it. The Organization section of Tempr's settings lists them: switch off a recommended one there; a required one stays on.
• Skills in your project. The agent finds skills in the standard SKILL.md format in .tempr/skills, .agents/skills and .claude/skills in your workspace, and in ~/.agents/skills, and reads one when a task matches it. A tool card shows "Skill: name" when it does.
• AGENTS.md and other agents' files. Workspace instructions now come from AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules and Cursor's rules as well as .github/copilot-instructions.md, as in the CLI, and are read again on every turn.

Added

• Tool Packs: see which MCP tools your organization lets you use. If your organization uses Tool Packs, the MCP servers settings list your packs and mark the servers (and tools) that aren't in them as "Not in your Tool Packs", and your own servers as "Not managed by your organization" when only its servers are allowed.

Changed

• A tool call your organization's Tool Packs don't include says so. The tool call shows "Not allowed by your organization" instead of a generic error, and the agent is told not to call it again.

Changed

• A remote MCP server that needs you to sign in shows a sign-in link. When a tool call needs you to sign in to its server first, the tool call shows "Sign-in needed" with a Sign in button, which opens the sign-in page in your browser, and a Retry button, which asks Tempr to try again once you've signed in. Before, the call just failed with an error.

Added

• Your organization's rules and skills. When your organization publishes rules and skills for its agents, the agent follows its rules on every turn, next to your project's own instructions, and reads one of its skills when a task matches it. The Organization section of Tempr's settings lists them: switch off a recommended one there; a required one stays on.
• Skills in your project. The agent finds skills in the standard SKILL.md format in .tempr/skills, .agents/skills and .claude/skills in your workspace, and in ~/.agents/skills, and reads one when a task matches it. A tool card shows "Skill: name" when it does.
• Slash commands. Type / in the chat for the commands your organization publishes, or Tempr's built-in ones, and your organization's personas replace the built-in ones.
• AGENTS.md and other agents' files. Workspace instructions now come from AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules and Cursor's rules as well as .github/copilot-instructions.md, as in the CLI.

Added

• Tool Packs. When your organization limits the MCP tools you can use with Tool Packs, a tool outside your packs now shows "Not allowed by your organization" on its card instead of a bare error, and the agent is told not to call it again. The MCP settings show which of each server's tools your packs include. Each run also tells Tempr which of your organization's MCP servers a tool came from, so the right pack applies even when two servers have similar names.

Added

• Signing in to a remote MCP server from the chat. When a remote MCP server needs you to sign in before its tool can run, the tool's card now shows a sign-in link and a Retry button instead of a bare error. Sign in opens the link in your browser; Retry asks the agent to try again.

Added

• Your organization's rules and skills. When your organization publishes rules and skills for its agents, the agent follows its rules on every turn, next to your project's own instructions, and reads one of its skills when a task matches it. The Organization section of Tempr's settings lists them: switch off a recommended one there; a required one stays on.
• Skills in your project. The agent finds skills in the standard SKILL.md format in .tempr/skills, .agents/skills and .claude/skills in your workspace, and in ~/.agents/skills, and reads one when a task matches it. A tool card shows "Skill: name" when it does.
• AGENTS.md and other agents' files. Workspace instructions now come from AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules and Cursor's rules as well as .github/copilot-instructions.md, as in the CLI.

Changed

• A tool your organization's Tool Packs leave out says so. When your organization limits which MCP tools its members can use and a call is refused, the tool shows "Not allowed by your organization" with "Your organization's Tool Packs don't include this tool.", instead of a bare error. The agent is told only that the tool isn't available, and doesn't call it again.
• The settings screen shows your organization's Tool Packs. The MCP servers list names the Tool Packs that apply to you, marks a team server none of them include as "Not in your Tool Packs", and, when your organization only allows its own servers, marks yours as "Not managed by your organization".
• Each run tells Tempr which of your organization's MCP servers its tools come from, so Tool Packs match every tool to the right server, even when two servers have similar names.

Changed

• A remote MCP server that needs you to sign in says so. When a tool's server asks you to sign in first, the tool shows "Sign-in needed" with a Sign in button that opens the sign-in page in your browser, and a Retry button that asks the agent to try again once you have, instead of a bare error. The agent is told only that you need to sign in, and doesn't call the tool again in that turn.

Fixed

• The settings and Browse MCP Servers screens are readable in light themes. Their cards, dropdowns and search boxes showed dark with dark text in any light theme; they now take your theme's colors.
• Text is easier to read in light and high-contrast themes. Faint labels, the green "Done" status, links and badges were too pale against a light background, and green on black was too dim in High Contrast.
• Your messages are readable in High Contrast Light. They showed white text on a white bubble.

Added

• Settings > General > Search the workspace on follow-ups, on by default. In Ask mode, Tempr searches your workspace for code related to every message. Turn it off to search only on a chat's first message; mentioning @workspace or #codebase still searches. Visual Studio and JetBrains IDEs have the same setting.

Changed

• Current screenshots on this page. The Marketplace and Open VSX listing shows the chat panel as it looks today, mid-turn with a proposed diff, and the current Portal dashboard and usage pages. Nothing in the extension itself changed from 1.1.12.

Added

• A dollar limit per agent turn. Settings > Agent limits has "Stop an agent turn after (US dollars)", blank by default. With an amount set, the server stops the turn before its next model call once its calls, sub-agents included, have cost that much, and the chat says what the turn spent against the limit; send another message to keep going. Your organization may set a lower limit. If a model's price isn't known, the chat says once that its calls count as free. A turn with a limit always runs on the server: if it can't start there, you see why, instead of the turn running without the limit.

Changed

• Paid MCP tools say so, once. When an MCP server's tool asks to be paid before it runs, the agent is told plainly that the tool needs payment, at what price when the server says, and that Tempr doesn't pay for tools, so it stops calling it. The tool's row in the chat shows "Paid tool, not run" with the price, and later calls are answered, until the MCP servers restart, without contacting the server again. A tool the server lists as paid is marked as paid for the agent before it's ever called, and a server that asks for payment just to connect shows that as its startup error. Nothing else from the payment request reaches the agent.

Fixed

• An MCP tool's error now reads as an error. A tool result the server flagged as an error came back to the agent as if the tool had worked. It now comes back as "Error: " and the tool's message, and the tool's row shows Error.
• A remote MCP server that accepts a connection but never answers no longer holds up the agent's start: it's given up on after 5 seconds, as intended. The time limit never applied before.

Fixed

• The tags on Browse MCP Servers cards ("official", "files" and so on) are readable in light themes. They had white text on a light gray background.
• Text is easier to read in light themes. Faint labels, the green "Done" status, links and badges were too pale against a light background.

Added

• Ask mode searches your project. Before answering in Ask mode, Tempr finds the code in your project most related to your message and gives it to the model, as in VS Code and Visual Studio, so questions about your codebase get answers about your code. Settings > General > "Search the workspace on follow-ups" (on by default) controls whether this happens on every message or only on a chat's first one; mentioning @workspace or #codebase always searches.

Fixed

• The chat's keyboard and screen reader support works. The script behind it never loaded in JetBrains IDEs. Now: - screen readers announce chat updates; - menus and pickers work with the arrow keys; - the history panel, license panel and settings dialogs keep focus inside while open and give it back when closed; - history entries can be opened from the keyboard.

Fixed

• No more "IDE error occurred" from Go to File. On recent IntelliJ-platform IDEs (seen on IntelliJ IDEA 2026.2), the first Go to File or Search Everywhere raised a red error balloon because the "Send to Tempr" intention had no description. It now has one, which also fills in its entry under Settings > Editor > Intentions > Tempr.

Added

• A dollar limit per agent turn. Settings has "Stop an agent turn after (US dollars)", blank by default. With an amount set, the server stops the turn before its next model call once its calls, sub-agents included, have cost that much, and the chat says what the turn spent against the limit. Your organization may set a lower limit. If a model's price isn't known, the chat says once that its calls count as free.

Changed

• Paid MCP tools say so, once. When an MCP server's tool asks to be paid before it runs (x402 or MPP), the agent is told plainly that the tool needs payment, at what price when the server says, and that Tempr doesn't pay for tools, so it stops calling it. The tool shows "Paid tool, not run" with the price, and later calls are answered without contacting the server again until the MCP servers restart. A tool the server lists as paid is marked as paid before it's ever called, and a server that asks for payment just to connect says so as its startup error. Nothing else from the payment request reaches the agent.

Fixed

• An MCP tool's error reads as an error. A tool that reported an error had its message passed to the agent as if the tool had worked.
• A turn that reaches its step limit says so in the chat, instead of stopping without a word.
• An MCP server that answers with an HTTP error, or with something that isn't JSON, gives that error instead of "(no response from MCP server)", and a server that fails its handshake is stopped rather than left running.

Changed

• Settings moved into the Tempr chat window, as in VS Code and JetBrains IDEs. The gear at the top of the chat window, or Tools > Tempr Settings, opens the same settings screen those IDEs have, with search, a Reset button on every changed setting, and an Account card with your license status and cloud sync. Changes apply as you make them; there's no OK or Apply. Tools > Options > Tempr now has one button that opens the screen. Everything you set in Tools > Options before is carried over the first time this version starts.
• Browse MCP Servers opens the same browse screen as in VS Code and JetBrains IDEs, instead of a separate window.
• The chat and the settings screen stay in step. Picking a model, an approval mode or which tools are on in the chat changes the setting, and changing it on the settings screen changes the chat.
• A call to a tool you've turned off is refused, even if a model asks for it anyway.

Fixed

• The chat follows your Visual Studio theme in Light. It used its own built-in light colors instead of the theme's, and faint labels, links and badges were too pale to read comfortably. Links are now darkened just enough to be readable when a theme's own link color isn't.

Added

• A dollar limit per agent turn. Options has "Stop an agent turn after (US dollars)", blank by default. With an amount set, the server stops the turn before its next model call once its calls, sub-agents included, have cost that much, and the chat says what the turn spent against the limit. Your organization may set a lower limit. A turn with a limit always runs on the server: if it can't start there, you see why, instead of the turn running without the limit.

Changed

• Paid MCP tools say so, once. When an MCP server's tool asks to be paid before it runs (x402 or MPP), the agent is told plainly that the tool needs payment, at what price when the server says, and that Tempr doesn't pay for tools, so it stops calling it. The tool shows "Paid tool, not run" with the price, and later calls are answered without contacting the server again until the MCP servers restart. A tool the server lists as paid is marked as paid before it's ever called, and a server that asks for payment just to connect says so as its startup error. Nothing else from the payment request reaches the agent.

Fixed

• An MCP tool's error reads as an error for MCP servers run as a local command. A tool that reported an error had its message passed to the agent as if the tool had worked.
• A refusal from Tempr's server shows the server's own message, instead of "Your license session has expired".
• A local MCP server that doesn't answer its handshake is stopped rather than left running.

Fixed

• Listing every file shows as "Listing files (all files)", as in VS Code and JetBrains, instead of "Listing files ()".

Fixed

• PowerShell script output comes back whole. run_powershell_script output was cut at 120 characters with "…", or at 80 on macOS and Linux, where pwsh also added colour codes, so the agent got paths and table values with their ends missing. It now comes back whole and plain, and Windows PowerShell's padding at the end of table rows is taken off.
• A failing PowerShell script's error comes back as text. run_powershell_script returned PowerShell's error messages as CLIXML (XML with escaped line breaks and colour codes, plus a progress record from Windows PowerShell), instead of the message itself.
• run_powershell_script on Windows takes scripts up to about 11,000 characters. It ran through cmd.exe, whose command lines stop at 8,191 characters, about 3,000 characters of script.

Fixed

• PowerShell script output comes back whole. run_powershell_script output was cut at 120 characters with "…", or at 80 on macOS and Linux, where pwsh also added colour codes, so the agent got paths and table values with their ends missing. It now comes back whole and plain, and Windows PowerShell's padding at the end of table rows is taken off.
• A failing PowerShell script's error comes back as text. run_powershell_script returned PowerShell's error messages as CLIXML (XML with escaped line breaks and colour codes, plus a progress record from Windows PowerShell), instead of the message itself.
• On Linux, run_powershell_script now finds an installed PowerShell. It always reported PowerShell as not installed there.

Fixed

• Command output comes back whole. PowerShell cut run_command and run_powershell_script output at 120 characters with "…", so the agent got paths and table values with their ends missing. It also padded every table row with spaces, which filled much of the 8,000 characters kept of a result. Both are fixed.
• A failing PowerShell script's error comes back as text. run_powershell_script returned Windows PowerShell's error messages as CLIXML, XML with escaped line breaks and a "Preparing modules for first use" progress record, instead of the message itself.
• A sub-agent's reply (from a delegated agent such as @review) is no longer saved into the conversation as an answer of its own, which the next turn then read as something Tempr had said.

Changed

• Symbol search finds the definition first. search_symbols lists where a name is declared, and as what, before the places it's used, in C#, TypeScript and JavaScript, Python, Go, Rust, Java, Kotlin, C and C++, Swift, Ruby and PHP, the same as the CLI. It no longer builds the embedding index first, so it answers at once and costs nothing, and a definition in a later file is no longer pushed out by the first few mentions.

Fixed

• The agent now sees a JavaScript monorepo's packages. Its file tools skipped every folder named packages (meant for NuGet's old packages folder), which hid a monorepo's whole source. That folder is now skipped only beside a .sln. Build output is left out where a build made it (Cargo's and Maven's target, Gradle's build, a package's dist and out) rather than every folder with those names, along with Python virtual environments and caches.
• Searching the code now covers Kotlin build scripts (.kts), Vue, Svelte and Astro components, .NET project files, and build files with no extension such as Makefile and Dockerfile.

Added

• Web search: on a Claude model through your Anthropic key, or an OpenAI model through your OpenAI key, the agent can now search the web with the provider's own search. Each search shows as "Searching the web for …", and the pages it finds are ones the agent can then open. The provider bills each search to your key ($10 per 1,000 on both), and the cost the chat shows includes it.
• Web search on Grok: xAI's models search the web with xAI's own search too, billed to your xAI key at $5 per 1,000 searches, where opening a page counts as one. The cost the chat shows includes it.

Changed

• Claude filters what it finds: on Claude Opus and Sonnet 4.6 and later, and Claude Fable, a web search now runs through code Claude writes and runs on Anthropic's side, which reads the results and keeps only what matters before it reaches the conversation. Anthropic made it for more accurate answers from fewer tokens. A search costs the same, and the code runs free.
• An agent turn on a Claude model caches its tool definitions apart from its instructions, so a call whose instructions changed (every turn's second call, which drops the project overview) reads the tools from the cache instead of paying to write them again.

Fixed

• A turn in which Claude searched a lot no longer stops halfway: when Anthropic pauses a search partway through, the agent carries on where it paused.

Changed

• Symbol search finds the definition first. search_symbols lists where a name is declared, and as what, before the places it's used, in C#, TypeScript and JavaScript, Python, Go, Rust, Java, Kotlin, C and C++, Swift, Ruby and PHP, the same as the CLI. It used to list the first matches in file order, so a definition in a later file could be cut off.

Fixed

• The agent now sees a JavaScript monorepo's packages. Its file tools skipped every folder named packages (meant for NuGet's old packages folder), which hid a monorepo's whole source. That folder is now skipped only beside a .sln. Build output is left out where a build made it (Cargo's and Maven's target, Gradle's build, a package's dist and out) rather than every folder with those names, along with Python virtual environments and caches.
• Searching the code now covers Kotlin build scripts (.kts), Vue, Svelte and Astro components, .NET project files, and build files with no extension such as Makefile and Dockerfile.

Added

• Web search: on a Claude model through your Anthropic key, or an OpenAI model through your OpenAI key, the agent can now search the web with the provider's own search. Each search shows as "Searching the web for …", and the pages it finds are ones the agent can then open. The provider bills each search to your key ($10 per 1,000 on both), and the cost the chat shows includes it.
• Web search on Grok: xAI's models search the web with xAI's own search too, billed to your xAI key at $5 per 1,000 searches, where opening a page counts as one. The cost the chat shows includes it.

Changed

• Claude filters what it finds: on Claude Opus and Sonnet 4.6 and later, and Claude Fable, a web search now runs through code Claude writes and runs on Anthropic's side, which reads the results and keeps only what matters before it reaches the conversation. Anthropic made it for more accurate answers from fewer tokens. A search costs the same, and the code runs free.
• An agent turn on a Claude model caches its tool definitions apart from its instructions, so a call whose instructions changed (every turn's second call, which drops the project overview) reads the tools from the cache instead of paying to write them again.

Fixed

• A turn in which Claude searched a lot no longer stops halfway: when Anthropic pauses a search partway through, the agent carries on where it paused.

Changed

• Symbol search finds the definition first. search_symbols lists where a name is declared, and as what, before the places it's used, in C#, TypeScript and JavaScript, Python, Go, Rust, Java, Kotlin, C and C++, Swift, Ruby and PHP, the same as the CLI. It no longer builds the embedding index first, so it answers at once and costs nothing, and a definition in a later file is no longer pushed out by the first few mentions.

Fixed

• The agent now sees a JavaScript monorepo's packages. Its file tools skipped every folder named packages (meant for NuGet's old packages folder), which hid a monorepo's whole source. That folder is now skipped only beside a .sln. Build output is left out where a build made it (Cargo's and Maven's target, Gradle's build, a package's dist and out) rather than every folder with those names, along with Python virtual environments and caches.
• Searching the code now covers Rust, Kotlin, Swift, Scala, TOML, Vue and Svelte files and build files with no extension such as Makefile and Dockerfile, which it used to skip.

Added

• Web search: on a Claude model through your Anthropic key, or an OpenAI model through your OpenAI key, the agent can now search the web with the provider's own search. Each search shows as "Searching the web for …", and the pages it finds are ones the agent can then open. The provider bills each search to your key ($10 per 1,000 on both), and the cost the chat shows includes it.
• Web search on Grok: xAI's models search the web with xAI's own search too, billed to your xAI key at $5 per 1,000 searches, where opening a page counts as one. The cost the chat shows includes it.

Changed

• Claude filters what it finds: on Claude Opus and Sonnet 4.6 and later, and Claude Fable, a web search now runs through code Claude writes and runs on Anthropic's side, which reads the results and keeps only what matters before it reaches the conversation. Anthropic made it for more accurate answers from fewer tokens. A search costs the same, and the code runs free.
• An agent turn on a Claude model caches its tool definitions apart from its instructions, so a call whose instructions changed (every turn's second call, which drops the project overview) reads the tools from the cache instead of paying to write them again.

Fixed

• A turn in which Claude searched a lot no longer stops halfway: when Anthropic pauses a search partway through, the agent carries on where it paused.

Changed

• Autopilot no longer runs every npm script. It used to run any package.json script unattended, npm run deploy included; now it runs npm test and the build, compile, test, lint and typecheck scripts, and asks before the rest. An allowed command covers whole words, so npm run build doesn't cover npm run build-and-deploy.
• Trusting a command for the session trusts that script, not the whole tool: trusting npm run build no longer trusts every npm run script, and python -m pytest no longer trusts every Python module.
• A build that also publishes (mvn test deploy, ./gradlew build publish) asks every time, like a delete.

Fixed

• The agent remembers its own answers. Each message used to reach it with your earlier questions in the chat but none of its replies, so a follow-up like "now make it faster" came without what "it" was. Agent and Ask mode both.

Added

• The agent asks before opening a page it found itself: without asking, it opens a link from your messages, from a tool's result or from a page it has already read, and any page of a well-known documentation site. For any other address it now shows a card with the whole address: Open, Trust the site for this session, or Reject, the way Claude Code asks per site. It used to refuse.
• Sites the agent can open: Settings > Permissions > Sites the agent can open lists sites whose pages open without asking: docs.example.com, or *.example.com for its subdomains. It stays on this machine, like your auto-approved commands.
• Links from earlier turns: a link on a page the agent read earlier in a chat can be opened in a later turn of the same chat.

Changed

• Page reads show as "Fetching <address>" and package lookups as "Looking up npm package react", instead of the tool's name.

Changed

• The agent can now open any page of a well-known documentation site, such as docs.python.org, developer.mozilla.org or learn.microsoft.com, without the link having appeared in the conversation first. Other addresses it didn't get from you, from a tool or from a page it read are still refused in this version.
• A web address that carries a key, token or password is opened only if you typed that credential yourself, and an address on your own machine or network (localhost, an intranet name) is answered with a pointer to curl instead of a failed request.

Fixed

• An agent turn on a model you reach through a provider's own API, such as Anthropic, OpenAI or DeepSeek directly, now shows what it cost. It showed $0 because only OpenRouter reports a price with each reply; the cost is now worked out from the model's price, the same figure the Portal's usage page always had.

Changed

• Autopilot no longer runs every npm script. It used to run any package.json script unattended, npm run deploy included; now it runs npm test and the build, compile, test, lint and typecheck scripts, and asks before the rest. An allowed command covers whole words, so npm run build doesn't cover npm run build-and-deploy.
• Trusting a command for the session trusts that script, not the whole tool: trusting npm run build no longer trusts every npm run script, and python -m pytest no longer trusts every Python module.
• A build that also publishes (mvn test deploy, ./gradlew build publish) asks every time, like a delete.

Fixed

• The agent remembers the conversation. Each message used to reach it as the start of a new chat, with none of your earlier questions or its answers.
• Workspace instructions and project context now reach the agent. It gets your .github/copilot-instructions.md with every message (Settings > General > Workspace instructions) and an overview of the project's files and README on each message's first step (Settings > Agent limits > Project context budget). Both settings were there but did nothing.

Added

• The agent asks before opening a page it found itself. Without asking, it opens a link from your messages, from a tool's result or from a page it has already read, and any page of a well-known documentation site. For any other address it now shows a card with the whole address: Open, Trust the site for this session, or Reject, the way Claude Code asks per site. It used to refuse.
• Sites the agent can open: Settings > Permissions > Sites the agent can open lists sites whose pages open without asking: docs.example.com, or *.example.com for its subdomains. It stays on this machine, like your auto-approved commands.
• Links from earlier turns: a link on a page the agent read earlier in a chat can be opened in a later turn of the same chat.

Changed

• Page reads show as "Fetching <address>" and package lookups as "Looking up npm package react", instead of the tool's name.

Changed

• The agent can now open any page of a well-known documentation site, such as docs.python.org, developer.mozilla.org or learn.microsoft.com, without the link having appeared in the conversation first. Other addresses it didn't get from you, from a tool or from a page it read are still refused in this version.
• A web address that carries a key, token or password is opened only if you typed that credential yourself, and an address on your own machine or network (localhost, an intranet name) is answered with a pointer to curl instead of a failed request.

Fixed

• An agent turn on a model you reach through a provider's own API, such as Anthropic, OpenAI or DeepSeek directly, now shows what it cost. It showed $0 because only OpenRouter reports a price with each reply; the cost is now worked out from the model's price, the same figure the Portal's usage page always had.

Changed

• Autopilot no longer runs every npm script. It used to run any package.json script unattended, npm run deploy included; now it runs npm test and the build, compile, test, lint and typecheck scripts, and asks before the rest. An allowed command covers whole words, so npm run build doesn't cover npm run build-and-deploy.
• Trusting a command for the session trusts that script, not the whole tool: trusting npm run build no longer trusts every npm run script, and python -m pytest no longer trusts every Python module.
• A build that also publishes (mvn test deploy, ./gradlew build publish) asks every time, like a delete.

Fixed

• The agent follows your .github/copilot-instructions.md for the whole chat. It was sent with a chat's first message only, so from the second message on the agent answered without it.

Added

• The agent asks before opening a page it found itself. Without asking, it opens a link from your messages, from a tool's result or from a page it has already read, and any page of a well-known documentation site. For any other address it now shows a card with the whole address: Open, Trust the site for this session, or Reject, the way Claude Code asks per site. It used to refuse.
• Sites the agent can open: Tools > Options > Tempr > Agent Commands > Sites the agent can open lists sites whose pages open without asking: docs.example.com, or *.example.com for its subdomains. It stays on this machine, like your auto-approved commands.
• Links from earlier turns: a link on a page the agent read earlier in a chat can be opened in a later turn of the same chat.

Changed

• Page reads show as "Fetching <address>" and package lookups as "Looking up npm package react", instead of the tool's name.

Fixed

• A finished page read or package lookup was remembered in the chat as "Delegated sub-agent completed". It's now recorded under its own name.

Changed

• The agent can now open any page of a well-known documentation site, such as docs.python.org, developer.mozilla.org or learn.microsoft.com, without the link having appeared in the conversation first. Other addresses it didn't get from you, from a tool or from a page it read are still refused in this version.
• A web address that carries a key, token or password is opened only if you typed that credential yourself, and an address on your own machine or network (localhost, an intranet name) is answered with a pointer to curl instead of a failed request.

Fixed

• An agent turn on a model you reach through a provider's own API, such as Anthropic, OpenAI or DeepSeek directly, now shows what it cost. It showed $0 because only OpenRouter reports a price with each reply; the cost is now worked out from the model's price, the same figure the Portal's usage page always had.

Added

• The agent can read web pages and look up packages: give it a link to a docs page, a GitHub file or a changelog and it reads the page as Markdown; ask about a library and it looks up the latest version, release date and links on npm, PyPI, crates.io, the Go module proxy, NuGet or Maven Central. Both run on Tempr's servers, so they work in this version with no update, and neither needs a key; the chat shows them as "Running web_fetch" and "Running lookup_package" until the next release names the page or package. The agent only opens addresses that appeared in your messages, in its tools' results, or as links on a page it has already read, so it can't make up an address that sends your code somewhere. Tempr's servers can't reach your own network or localhost.

Added

• The agent can read web pages and look up packages. Give it a link to a docs page, a GitHub file or a changelog and it reads the page as Markdown; ask about a library and it looks up the latest version, release date and links on npm, PyPI, crates.io, the Go module proxy, NuGet or Maven Central. Both run on Tempr's servers, so they work in this version with no update, and neither needs a key; the chat shows them as "Running web_fetch" and "Running lookup_package" until the next release names the page or package. The agent only opens addresses that appeared in your messages, in its tools' results, or as links on a page it has already read, so it can't make up an address that sends your code somewhere. Tempr's servers can't reach your own network or localhost.

Added

• The agent can read web pages and look up packages. Give it a link to a docs page, a GitHub file or a changelog and it reads the page as Markdown; ask about a library and it looks up the latest version, release date and links on npm, PyPI, crates.io, the Go module proxy, NuGet or Maven Central. Both run on Tempr's servers, so they work in this version with no update, and neither needs a key; the chat shows them by their tool names until the next release names the page or package. The agent only opens addresses that appeared in your messages, in its tools' results, or as links on a page it has already read, so it can't make up an address that sends your code somewhere. Tempr's servers can't reach your own network or localhost.

Fixed

• On Visual Studio 2022 and early Visual Studio 2026 builds, signing in could fail with an assembly version error ("The located assembly's manifest definition does not match the assembly reference"). Every assembly the extension ships with is now registered, so Visual Studio loads the right one.

Fixed

• On Visual Studio 2022 and early Visual Studio 2026 builds, View > Tempr did nothing and Tools > Options > Tempr > General showed "An error occurred loading this property page". The extension now ships the System.Text.Json version it needs, with its dependencies.

Changed

• Perplexity's Sonar models are now its Agent API presets, perplexity/fast, low, medium, high and xhigh, since Perplexity retires Sonar chat completions on 2026-09-27. A Sonar model you had picked moves to its replacement by itself, so it stays selected: sonar becomes fast, sonar-pro becomes low, sonar-reasoning-pro becomes medium, and sonar-deep-research becomes high. That covers your chat model, inline-completion model, per-agent models and each saved chat's model.

Fixed

• The extension's Marketplace description now says what it does with your data: requests go through Tempr's server to the model you picked, on your own provider key, and chat history syncs to your Tempr account unless you turn Cloud sync off. It used to say chat history stayed on your machine. It also lists Teams plans as available, from $35 per seat per month (10-seat minimum), instead of coming soon.

Changed

• Perplexity's Sonar models are now its Agent API presets, perplexity/fast, low, medium, high and xhigh, since Perplexity retires Sonar chat completions on 2026-09-27. A Sonar model you had picked moves to its replacement by itself, so it stays selected: sonar becomes fast, sonar-pro becomes low, sonar-reasoning-pro becomes medium, and sonar-deep-research becomes high. That covers your chat model, inline-completion model and per-agent models.

Fixed

• The plugin's Marketplace description now says what it does with your data: requests go through Tempr's server to the model you picked, on your own provider key; chat history syncs to your Tempr account unless you turn Cloud sync off; and error reports are sent automatically unless you turn them off. It used to say chat history stayed on your machine.

Added

• Cloud sync switch: Tempr Settings now has a Cloud sync switch in the Account card. Turn it off and Tempr stops storing your chats, custom agents and tool preferences on your Tempr account; your chats stay on this machine. The switch belongs to your account, so it covers every machine and every Tempr extension signed in to it, and you can also change it in the Tempr Portal. Turning it off offers to delete what's already stored, and Delete synced data does that at any time. Sync stays on unless you turn it off.

Changed

• Perplexity's Sonar models are now its Agent API presets, perplexity/fast, low, medium, high and xhigh, since Perplexity retires Sonar chat completions on 2026-09-27. A Sonar model you had picked moves to its replacement on its own, so it stays selected: sonar becomes fast, sonar-pro becomes low, sonar-reasoning-pro becomes medium, and sonar-deep-research becomes high. That covers your chat model, inline-completion model and per-persona models.

Added

• Cloud sync switch: Tools > Options > Tempr > General has a new setting, Sync chats and settings to your Tempr account. Turn it off and Tempr stops storing your chats, tool preferences, custom agents and code search index on your Tempr account. Your chats stay on this machine. The setting belongs to your account, so it covers every machine and every Tempr extension signed in to it, and it takes effect as soon as you change it. When you turn it off, Tempr asks whether to also delete what's already stored. Delete synced data... does that at any time and ends any share links. Sync stays on unless you turn it off.

Fixed

• Your solution's code search index never reached your Tempr account: every upload failed without a message. Uploads now work, and each solution's existing local index is uploaded once. Until the synced index has caught up, search keeps ranking with the index on your machine, so results never get worse while it fills.

Added

• Cloud sync switch. Tempr: Open Settings now has a Cloud sync switch in the Account card. Turn it off and Tempr stops storing your chats, custom agents and tool preferences on your Tempr account; your chats stay on this machine. The switch belongs to your account, so it covers every machine and every Tempr extension signed in to it, and you can also change it in the Tempr Portal. Turning it off offers to delete what's already stored, and Delete synced data does that at any time. Sync stays on unless you turn it off.

Fixed

• If your IDE can't run its built-in browser (JCEF), for example because it was started on a Java runtime without it, Tempr no longer raises an IDE error every time a project opens. The Tempr panel instead explains why the chat can't start, with a link to Choose Boot Java Runtime for the IDE so you can switch to a runtime that includes JCEF.

Security

• Automatic error reports no longer include your license key. A report is now linked to your account by your signed-in session. It carries only that session's activation ID, the error message and the stack trace.

Changed

• Google's Gemini models are now listed as google/… instead of gemini/…, the same name OpenRouter, Vercel's AI Gateway and models.dev give them, so a model id copied from any of those works in Tempr as it is. Nothing you have set needs changing: your chat model, inline-completion model, per-agent models and reasoning levels follow the new name by themselves, and the old gemini/… name keeps working everywhere.

Changed

• Google's Gemini models are now listed as google/… instead of gemini/…. That's the name OpenRouter, Vercel's AI Gateway and models.dev give them, so a model ID copied from any of those works in Tempr as it is. You don't need to change anything: your chat model, inline-completion model, per-persona models and reasoning levels move to the new name on their own, and the old gemini/… name still works everywhere.

Changed

• Google's Gemini models are now listed as google/… instead of gemini/…, the same name OpenRouter, Vercel's AI Gateway and models.dev give them, so a model id copied from any of those works in Tempr as it is. Nothing you have set needs changing: your chat model, inline-completion model, per-agent models and reasoning levels follow the new name by themselves, and the old gemini/… name keeps working everywhere.

Added

• Reasoning level. A Reasoning dropdown beside the model dropdown sets how much the model thinks before it answers: Default (the model's own setting), Off, or the levels that model offers, from Minimal to Max. Tempr remembers your pick for each model, and it applies when you use a Tempr model from Copilot Chat too. The dropdown only appears for models that have reasoning control, and a level a model lacks is sent as the nearest one it has.

Fixed

• After you first signed in, the model dropdown could stay empty for up to five minutes. It now fills in right away.

Added

• Reasoning level. A Reasoning dropdown beside the model dropdown sets how much the model thinks before it answers: Default (the model's own setting), Off, or the levels that model offers, from Minimal to Max. Tempr remembers your pick for each model. The dropdown only appears for models that have reasoning control, and a level a model lacks is sent as the nearest one it has.

Fixed

• The Tempr tool window's sidebar icon was drawn at the full size of the logo. It's now the same size as the other tool window icons, in both the classic and the new UI.

Added

• Reasoning level: a new Reasoning dropdown beside the model dropdown sets how much the model thinks before it answers. The choices are Default (the model's own setting), Off, and whichever levels that model offers, from Minimal to Max. Tempr remembers your choice for each model. The dropdown only appears for models that support reasoning control, and a level a model doesn't have is sent as the nearest one it does.

Added

• Settings follow you with Settings Sync. With VS Code Settings Sync turned on, your Tempr settings reach your other machines: the chat model, approval mode, inline completions, agent limits, chat font size and the rest. MCP servers and auto-approved commands stay on each machine, and custom agents keep syncing through your Tempr account.

Security

• A repository's .vscode/settings.json can no longer point Tempr at a different server. The development override moved from the tempr.serverUrl setting to the TEMPRAPP_SERVER_URL environment variable, the same one Visual Studio reads.
• A repository's .vscode/settings.json can no longer change Tempr's settings, such as switching approval mode to Autopilot, pre-approving commands, or adding MCP servers or custom agents. Tempr now reads tempr.* values only from your user settings.
• A symbolic link or Windows junction inside your workspace can no longer lead Tempr's file tools outside it. Reading, writing, patching and searching files now follow a link to where it really goes, and refuse it when that's outside the workspace.
• Every spelling of a delete command is now treated as destructive, so it asks for approval every time, in every approval mode, and can't be trusted for the session. Before, rm -fr and Remove-Item -Force -Recurse weren't recognized, so approving rm -fr build once trusted rm -fr on any path for the rest of the session. Tempr now goes by the command itself (rm, rmdir, rd, del, erase, Remove-Item, ri, unlink), whatever its flags or a sudo in front.

Security

• A symbolic link or Windows junction inside your project can no longer lead Tempr's file tools outside it. Reading, writing, patching, listing and searching files now follow a link to where it really goes, and refuse it when that's outside the project.
• Every spelling of a delete command is now treated as destructive, so it asks for approval every time, in every approval mode, and can't be trusted for the session. Before, rm -fr and Remove-Item -Force -Recurse weren't recognized, so approving rm -fr build once trusted rm -fr on any path for the rest of the session. Tempr now goes by the command itself (rm, rmdir, rd, del, erase, Remove-Item, ri, unlink), whatever its flags or a sudo in front.

Security

• A symbolic link or Windows junction inside your solution can no longer lead Tempr's file tools outside it. Reading, writing, patching, listing and searching files now follow a link to where it really goes, and refuse it if that's outside the solution. A link that points back at one of its own parent folders no longer sends a search into a loop.
• Every spelling of a delete command now counts as destructive, so it asks for approval every time, in every approval mode, and can't be trusted for the session. Before, rm -fr and Remove-Item -Force -Recurse weren't recognized, so approving rm -fr build once trusted rm -fr on any path for the rest of the session. Tempr now goes by the command itself (rm, rmdir, rd, del, erase, Remove-Item, ri, unlink), whatever its flags and even with sudo in front.

Fixed

• Agent turns no longer lose their connection when the chat goes quiet for more than ten seconds, such as while a model thinks or a tool waits for your confirmation. Before, a long enough quiet stretch could end the turn with "Lost connection to the agent run".

Added

• Accessibility: screen readers now hear when a reply finishes, when it fails, and when a tool call is waiting for confirmation. Chat tabs, the mode, approval mode, and model dropdowns, the attach menu, and chat history all work from the keyboard. The history, license, and tools panels and the settings dialogs behave as dialogs, and keyboard focus is always visible.
• Vault MCP servers in settings: MCP servers assigned to your account in Tempr's vault now appear in the Tempr Settings MCP list, as read-only cards.
• Sub-agent cards: when the agent delegates work to a sub-agent on the server, the delegation now shows as a tool card in the chat, with the sub-agent's messages under it.

Changed

• If the Tempr server speaks a newer agent-run protocol than this version understands, the plugin says so once per session instead of failing in unclear ways.
• Editing a message you've already sent is hidden for now: saving an edit never worked in JetBrains IDEs, and it comes back once it does.

Changed

• If the Tempr server uses a newer agent protocol than your version of the extension understands, Tempr tells you so instead of failing in unclear ways.
• The extension's Marketplace description no longer mentions inline chat, which is turned off for now.

Fixed

• When the agent ran on the Tempr server, a tool you approved and Visual Studio ran could have its result replaced by the server's generic summary. For example, "Read 12 lines." became "Done", or a rejected call showed as an error. Results now keep their real summaries.

Fixed

• MCP tool calls to Anthropic models could fail outright: qualified MCP tool names ({server}.{tool}) contained a ., which Anthropic's API rejects in a tool name. Qualified names now use {server}__{tool} instead, matching the Tempr CLI and the Visual Studio and JetBrains extensions.

Fixed

• MCP tool calls to Anthropic models could fail outright: qualified MCP tool names ({server}.{tool}) contained a ., which Anthropic's API rejects in a tool name. Qualified names now use {server}__{tool} instead, the same as Tempr's other extensions and the CLI.
• An agent turn no longer fails when the connection drops or stalls partway through. Tempr reconnects and picks the turn up where it left off, without repeating anything you've already seen, and only gives up after several attempts.

Fixed

• MCP tool calls to Anthropic models could fail outright. MCP tool names were shown to the model as {server}.{tool}, and Anthropic's API doesn't allow a . in a tool name. They are now {server}__{tool}, the same as in Tempr's other extensions.

Added

• Vault MCP servers: MCP servers assigned to your account in Tempr's server-side vault now show up in the extension automatically, merged in with your local tempr.mcpServers entries (a vault entry wins on a name collision, and vault servers appear as read-only cards in the Tempr Settings MCP list).
• Remote MCP servers: vault-assigned remote servers connect over MCP's Streamable HTTP transport through Tempr's own proxy relay instead of to the remote server directly, so the real remote URL and its credentials never leave the Tempr server.

Fixed

• A vault-fetch failure or a slow/hanging remote MCP connect (now capped at 5 seconds) no longer delays or blocks local MCP servers from starting.

Added

• Vault MCP servers: MCP servers assigned to your account in Tempr's server-side vault now start alongside the ones you've configured yourself, and their tools are available to the agent. When a vault server and one of yours share a name, the vault's entry wins.
• Remote MCP servers: vault-assigned remote servers connect over MCP's Streamable HTTP transport through Tempr's own relay instead of to the remote server directly, so the real remote URL and its credentials never leave the Tempr server.

Changed

• The plugin is now called Tempr Chat on the JetBrains Marketplace and in the IDE's plugin list.

Fixed

• A vault fetch that fails or hangs (now capped at five seconds) no longer delays or blocks your own MCP servers from starting.
• Works with newer IDE versions, which deprecated or changed some of the platform APIs the plugin used, including the one that serves the chat panel into the IDE's built-in browser.

Added

• Vault MCP servers: MCP servers assigned to you in Tempr's server-side vault, your own or shared by your team, now show up in the extension automatically alongside the servers you set up yourself. Tools > Options > Tempr > MCP Servers lists them in a read-only Vault-managed servers table. You manage them in the Tempr Portal.
• Remote MCP servers: remote servers from the vault connect through Tempr's own relay, not to the remote server directly, so the real remote URL and its credentials never leave the Tempr server. You need to be signed in to use them. A remote server that can't connect is skipped, and your other servers keep working.

Fixed

• Chat scrollbar: the message list's scrollbar (and its clickable/hoverable area) now stops above the input bar instead of running down behind it.
• Chat input box: after switching chat tabs, the input box's auto-resize height sometimes computed too tall while its layout was still settling; it now recalculates once the box's width has actually settled.

Fixed

• The tools/skills panel's "Select all" is now a single panel-level toggle that flips to "Deselect all" once every tool is enabled, replacing two redundant "Select all" links (one panel-level, one scoped to the "Built-in" section).
• Settings key-value editors (custom headers, environment variables, etc.) now drop a removed row from view immediately when its Remove button is clicked, instead of leaving it on screen until it lost focus.

Changed

• In the tools picker, the button that switches every tool on or off now says Deselect all once every tool is on. Each group's button is now called Toggle all.

Fixed

• On plans that include delegation to sub-agents, every Agent-mode and persona turn could fail with a "Tool names must be unique" error from the model provider. Those turns work again.

Changed

• Tiered delegation concurrency. How many delegated sub-agents can run at the same time now depends on your plan: 2 on Standard and 8 on Pro. Before, every plan could run 8. When the agent asks for more at once, the extra sub-agents wait their turn.

Changed

• Tiered delegation concurrency. The cap on concurrently-running delegated sub-agents ("delegate_to_agent") is now tier-based (Standard: 2, Pro/Team: 8) instead of a flat 8, matching the JetBrains/Visual Studio entitlement limits. The cap is re-applied on every delegation, so a sign-in, sign-out, or plan upgrade mid-session takes effect immediately.

Added

• Server-side agent loop: agent tool-calling turns can now be driven by the Tempr server, with MCP tool catalogs resolved server-side. Tool results are delivered with automatic retries, so a transient network hiccup no longer leaves a chat frozen mid-turn.
• Session-scoped tool trust: confirmation cards for shell commands offer "Trust <prefix> for this session" (e.g. npm run), and PowerShell script approvals can be remembered per tool. Session prefixes match on exact command tokens, so trusting npm run never accidentally trusts npm runbook.
• The responding indicator now stays pinned under the streaming reply with a live elapsed-time and token-estimate ticker, instead of disappearing at the first token.

Fixed

• Assorted chat-window rendering issues and markdown formatting fixes.

Added

• Server-side agent loop: agent turns can now run on the Tempr server, and from 2.2.10 so can Ask-mode turns. Tools still run in Visual Studio, with your approval where it's needed. Tool results are sent with automatic retries, so a brief network problem no longer leaves a chat stuck partway through a turn.
• Browse MCP Servers: a new plug button in the chat toolbar opens a catalog of MCP servers you can install in one click. If a server needs environment variables, such as an API key, Tempr asks for them. A newly installed server starts on your next agent turn, with no restart needed. Before, a server you added didn't start until Visual Studio restarted.

Changed

• Between a tool call and the model's next reply, the status line under the transcript now says "Working…" instead of disappearing, so a long turn no longer looks frozen.

Fixed

• A scrollbar no longer flashes in and out each time you send a message.

Changed

• The Tools > Options > Tempr pages for allowed command prefixes, custom agents, MCP servers and per-persona models are now editable tables with Add and Remove buttons. They replace the pipe-delimited text boxes. A custom agent's system prompt opens in its own editor, and per-persona model overrides offer a dropdown of your models.

Added

• Chat history across machines: chats saved from another machine signed in to your license now download to this one. A fresh install starts with your history instead of an empty list.

Changed

• Chat font size moved from the chat panel to Tools > Options > Tempr > General. It applies to open chats right away, and Ctrl+=, Ctrl+- and Ctrl+0 still adjust it in the chat window.
• The chat toolbar's gear button now opens Tools > Options > Tempr. A new key button opens your license key and a link to manage it in the Tempr Portal.
• The header above each chat has been removed. Its message count now appears at the start of the toolbar's usage line.

Fixed

• Reopening a chat from history no longer leaves a large blank gap above the last message.

Changed

• The Feature/Bug request button now opens the feature request page in the Tempr Portal instead of a form inside Visual Studio.

Fixed

• The agent's run_tests tool always failed before it ran any tests. It now runs your tests and reports the results.

Fixed

• A chat's title now comes from the question you typed. Before, it could come from code attached to the message, such as a snippet sent with Send to Tempr.