MCP servers in the AI agent
MCP (Model Context Protocol) servers give the agent extra tools. The agent supports two kinds:
| Kind | Where it runs | Where it is configured |
|---|---|---|
| Hosted | Outside your machine; CHAMPREP connects to it over HTTPS | Your CHAMPREP AI account, shared with the web, desktop, and mobile apps |
| Local | A program started on your machine, speaking MCP over standard input and output | An mcp.json file for you or for the project |
This page is about connecting servers to the agent. To use CHAMPREP itself from another MCP client, see the CHAMPREP MCP server.
The /mcp window
Section titled “The /mcp window”/mcpThe window lists every hosted and local server with its status and tools. Select a server with the arrow keys and press:
| Key | Action |
|---|---|
Enter | Show the actions for the server. |
e | Turn the server on or off. |
r | Refresh: reconnect and reload its tools. |
a | Re-authenticate: replace its credential. |
t | List its tools. |
x | Remove the server. |
n | Add a server. |
Shortcuts also work without the window, for example /mcp list,
/mcp enable NAME, /mcp disable NAME, /mcp refresh NAME, and, for local
servers, /mcp restart NAME and /mcp remove NAME.
Hosted servers
Section titled “Hosted servers”Hosted servers belong to your CHAMPREP AI account, so a server you add in the terminal is also available on the web, and the other way round. Their tools run on CHAMPREP, not on your machine.
To add one, press n in /mcp, choose a hosted server, and enter its URL and
credential, such as a bearer token. The credential is entered in a masked
field, sent once to your CHAMPREP AI account, and never shown, saved locally,
logged, or given to the model. Use a to replace an expired credential.
Removing a hosted server removes it from your account on every surface. If a hosted-server action is not available from the CLI in your environment, the agent says so; use your CHAMPREP AI settings on the web instead.
Local servers
Section titled “Local servers”Local servers are programs the agent starts on your machine. Add one with n
in /mcp and choose:
- the server name and the command with its arguments;
- for just you (your user
mcp.json) or for this project (.champrep/mcp.json, which you can commit); - optionally, an environment variable that holds a credential. Its value is
masked, stored only in your user
mcp.json, and never written to a project file, history, logs, or the model’s context.
You can also edit the files directly. They use the common MCP client format. Keep credentials in your user file, never in a project file you commit:
{ "mcpServers": { "github": { "command": "github-mcp-server", "args": ["stdio"], "env": { "GITHUB_TOKEN": "YOUR_TOKEN" } } }}Optional fields per server: disabled, include_tools, exclude_tools,
startup_timeout_sec (default 30), tool_timeout_sec (default 120, maximum
600), and priority. Only local (stdio) servers belong in these files; add
HTTP servers as hosted servers.
Local servers never receive your CHAMPREP credentials or any CHAMPREP_*
environment variable.
Run champrep ai paths to find your user mcp.json. Local MCP servers are a
plan feature; see Plans and limits.
Project servers and trust
Section titled “Project servers and trust”A project’s .champrep/mcp.json runs commands on your machine, so its servers
start only after you trust the project’s .champrep/ configuration. If the
file changes, you are asked again. See
Repository trust.
Approving MCP tool calls
Section titled “Approving MCP tool calls”- Local server tools are treated like commands: each call asks for your
permission, showing the server and arguments, unless an allow rule covers
it. For example,
local_mcp__github__*allows every tool of thegithubserver. Deny rules work the same way. - Hosted server tools follow the approval setting of the server in your CHAMPREP AI account. When a call needs approval, the agent shows an approval card; nothing is approved automatically.
The agent can offer at most 64 local tools per turn. When local servers
provide more, lower-priority servers are trimmed first and the agent tells
you. Use include_tools or exclude_tools to keep the tools you need.