MCP servers
Last updated
Model Context Protocol (MCP) is an open standard for connecting an AI agent to external tools and data sources. Tempr's agent can use any MCP server you add, right alongside its built-in tools.
Browsing the catalog
Open the MCP panel inside the chat tool window to search and one-click install curated servers without leaving your IDE — the same catalog shown on the marketing site's MCP catalog. A few of the official servers you'll find there:
| Server | Publisher | What it does |
|---|---|---|
| Filesystem | Anthropic | Read and write files/directories on the local filesystem, scoped to a directory you choose. |
| Fetch | Anthropic | Fetch and convert web pages to Markdown for the model to read. |
| Memory | Anthropic | A persistent knowledge-graph memory the agent can read and write across a session. |
| GitHub | Official | Issues, pull requests, repos, and code search against GitHub. |
Adding a custom server
Not every server needs to come from the catalog. Add any MCP-compliant server manually with its command/args (for local stdio servers) or URL (for remote HTTP/SSE servers), and the agent will discover its tools automatically the next time it starts a turn.
Remote servers through Tempr
A remote server you add in the Portal under MCP credentials is reached through Tempr's relay, from every Tempr client: VS Code, JetBrains, Visual Studio, the CLI and Tempr Code. The server's URL and secrets stay on Tempr's servers, so laptops and CI runners never hold them, and every call shows up in your logs. Any other MCP client can use the same relay with a virtual key; see the remote MCP proxy.
Signing in to a server
A remote server can sign each person in with their own account (OAuth), so the agent acts as you, with your permissions, rather than as a shared bot. When a server needs you to sign in, the first tool call shows Sign in to {server} to use this tool with a Sign in button that opens the Portal, and a Retry button that asks the agent to try again once you're done. The CLI prints the link instead. You can see and sign out of every server under MCP sign-ins in the Portal. See per-member sign-in for how admins set it up.
Tools your organization allows
An organization can give each person only the MCP tools their job needs, with Tool Packs. A tool outside yours is left out of the agent's tool list, and refused if it's called anyway. When you need one, ask for it: the refusal links to a Portal page that shows the tool and asks why you need it, and your organization's owners and admins get an email. They can approve it for an hour, a day, a week or for good, or decline it with a note, and you get an email either way. An approval works from your next agent turn. See asking for a tool.
Guardrails on tool calls
The relay checks tool calls on their way out and results on their way back: secrets and personal data are flagged or redacted, text that looks like instructions planted for the model is flagged, and hidden characters are removed. Admins choose how strict each check is. See MCP guardrails. Before you rely on them, paste a tool call, result or description into the guardrail tester in the Portal to see what each check would do; the text isn't kept.
Where MCP applies
MCP tool execution is part of the agent loop: it runs alongside file edits and shell commands in your IDE or terminal session. Local (stdio) servers run on your machine; remote servers go through the relay above. If you're building your own agent against the Gateway API, you own the tool-calling loop, and can use the relay for remote MCP servers or wire up tools yourself with the tools parameter.