Pinecall

The MCP Server

Pinecall as an MCP server — build, run and debug voice agents from Claude Code, Cursor, or any MCP client.

read these docs, discover the model/voice/phone catalog, configure a dev agent, iterate on it by chatting with it, wire a phone number, and read the live call log of a real call — without leaving the editor.

Install#

pinecall mcp install

One command, every IDE — it detects the assistants installed on the machine (Claude Code, Codex, Antigravity, Cursor, Windsurf, Gemini CLI) and writes the pinecall entry into each one's own config. Re-running repairs a drifted entry; --remove uninstalls. Or by hand, in any MCP client:

{ "mcpServers": { "pinecall": { "command": "npx", "args": ["-y", "@pinecall/mcp"] } } }

Auth is PINECALL_API_KEY — from the environment, ~/.pinecall/credentials, or a read-only scan of your shell rc files (IDEs launch MCP servers without your shell; the server finds the key anyway and persists it so the scan runs once).

What it can do#

Eighteen tools, in journey order: whoami · set_api_key · docs_search · get_doc · knowledge · list_models · list_voices · play_voice · list_phones · list_agents · configure_agent · run_agent · chat · observe · list_calls · get_call · subscribe · byok.

Three rules shape it:

  • dev- only. configure_agent refuses any slug that does not start with dev-, with no override: registering an agent hot-reloads the live one, so a production slug would clobber the process that owns it. Production agents are deployed from code.
  • Chat is the testing story. There is deliberately no spec-runner tool — converse with the agent and judge the transcript. (pinecall test still exists in the CLI for YAML specs.)
  • Real code tools need a process. A tool() with an execute function cannot travel over MCP; run_agent runs your own agent file through the same machinery as pinecall run, with start/stop/status/logs.

The full tool reference, the safety model and the development setup live in the package's own repo: github.com/pinecall/mcp.

What's next#