AI agents

Kimi Code CLI

MCP: Connecting Servers to Kimi Code

Kimi Code CLI is an MCP client. MCP tools appear alongside built-in tools with no behavioral difference, named mcp__<server>__<tool> (e.g. mcp__github__create_issue).

The short version

  1. In the TUI, run /mcp-config and add the server interactively (it also handles MCP OAuth login via /mcp-config login <server-name>).
  2. Or edit mcp.json directly (below).
  3. Run /mcp to confirm the server connected.

mcp.json

Config locations: user-level ~/.kimi-code/mcp.json (or $KIMI_CODE_HOME/mcp.json) and project-level .kimi-code/mcp.json. On same-name entries, project wins. MCP servers are not configured in config.toml.

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    },
    "linear": {
      "url": "https://mcp.linear.app/mcp"
    },
    "legacy-events": {
      "transport": "sse",
      "url": "https://mcp.example.com/sse"
    }
  }
}

Transports: stdio (child process, via command/args), HTTP (via url), and sse (legacy; explicit transport: "sse").

Optional per-server fields: env, cwd (stdio), headers, bearerTokenEnvVar (HTTP/SSE), enabled, startupTimeoutMs (default 30000), toolTimeoutMs (default 60000), enabledTools, disabledTools. Timeout precedence: per-server > env var (KIMI_MCP_STARTUP_TIMEOUT_MS / KIMI_MCP_TOOL_TIMEOUT_MS) > config.toml [mcp] > built-in default.

A concrete example: GitHub over stdio

# 1. Add to ~/.kimi-code/mcp.json:
{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "ghp_…" }
    }
  }
}
# 2. In the TUI, verify:
/mcp

# 3. Use it — the agent now has tools like mcp__github__create_issue:
"Create a GitHub issue for the flaky test in src/api.test.ts"

Things to know

  • Trust: project-level MCP servers in an untrusted folder appear in the workspace trust prompt (default answer: “Trust this folder”). Review the launch command/URL before confirming — trusting enables the project’s MCP servers.
  • Mid-session changes: deleting a server doesn’t kill open sessions (listed as removed; calls fail); servers added mid-session only join new sessions.
  • Permissions: rules support */** wildcards (mcp__github__*); MCP tool args are not part of matching. Unmatched calls trigger approval. In Ask When Needed mode MCP calls are auto-approved — the docs warn to use that mode only with trusted servers. See Permissions and Approvals.
  • Plugins can declare mcpServers in their manifest (enabled by default; toggle with /plugins mcp enable|disable <id> <server> then /reload).

Last verified: 2026-09-07

Official sources:

AI coding tools evolve quickly. If you find outdated information, please open an issue or pull request.