pincerhq/pincer/mcp-server-setup

MCP server setup

Pincer is both an MCP client (connects to external MCP servers like GitHub, Slack, Postgres, filesystem, browser automation) and an MCP server (exposes Pincer's own tools to external MCP clients via OAuth 2.0). Skills and MCP servers now coexist — connecting an MCP server does not disable skills.

Connecting an external MCP server (client mode)

1. Install the optional extra: uv pip install "pincer-agent[mcp]".
2. Add a [[mcp.servers]] block to pincer.toml (or the equivalent env

vars) with the server's name, transport (stdio/SSE/HTTP), command or URL, and any required auth (API key, OAuth token).

1. Restart the agent (pincer run). On startup you'll see

MCP '<name>' connected or a failed to connect warning in the console — check that message first if something's wrong.

1. MCP tools appear in the agent's tool list identically to built-in

tools; the LLM calls them the same way, with the same approval flow.

For the full config reference (transports, auth schemes, scopes, examples for popular servers), load the detailed guide at docs/guides/mcp-guide.md in the repository, or reference it via load_skill_reference if it's bundled alongside this skill.

Exposing Pincer as an MCP server

Pincer's own tools can be exposed to external MCP clients (Claude Desktop, other agents) via OAuth 2.0 with PKCE and scope-based access control. Which tools are exposed by default is controlled by an expose_tools allowlist in the MCP config — read-only tools are typically exposed by default, execute/write tools are opt-in.

Troubleshooting

• "MCP config load failed" at startup: check pincer.toml syntax —

the agent falls back to running without MCP rather than crashing.

• A server shows connected but its tools aren't callable: check the

server's own scope/permission settings, not just Pincer's config.

• Too many tools registered (>100 from one server): the agent warns

that LLM tool selection may degrade — consider filtering which tools that server exposes.