@clawcall/mcp

MCP server wrapping the ClawCall REST API for AI-powered US phone calls.

ClawCall places real outbound phone calls in the US (+1 only). This package is a stdio MCP server that exposes the ClawCall REST contract as tools any MCP host (Claude Code, Claude Desktop, Cursor, Windsurf) can call. It is a stateless client over the public API at https://api.clawcall.dev — it imports nothing from the ClawCall server.

Install / register

Add the server to your MCP client config:

{ "mcpServers": { "clawcall": { "command": "npx", "args": ["-y", "@clawcall/mcp"], "env": { "CLAWCALL_API_KEY": "optional", "CLAWCALL_BASE_URL": "https://api.clawcall.dev" } } } }

Or, for Claude Code, the one-liner:

claude mcp add clawcall -- npx -y @clawcall/mcp

The env block is optional. The ClawCall server self-provisions an API key on the first call: an anonymous place_call auto-issues a proto-key, which this server captures and persists. You do not need to set CLAWCALL_API_KEY to get started.

Identity sharing

This MCP server reads and writes the same ~/.config/clawcall/key.json file ({ "api_key": "...", "user_phone_number": "..." }) that the ClawCall agent skill uses. Running both the skill and this MCP server does not fork your identity — they share one key, one trial quota, one call history, and one balance.

If CLAWCALL_API_KEY is set in the environment, it shadows the file (an operator / CI override). The saved file is otherwise the source of truth, and the proto-key issued on your first call is persisted there automatically (preserving any saved user_phone_number).

The workflow

A call runs for several minutes. The load-bearing loop is:

1. place_call — returns a call_id immediately. Does not block.
2. Poll get_call every 3 seconds until lifecycle === "finalized".
3. outcome is the phone-network result (answered / no_answer / busy / …), NOT task success. An answered call can still fail the task.
4. Read the transcript with get_call_transcript and judge it against the task you set.

Choosing a tool

• place_call_and_wait is the single-foreground-call shortcut: it places the call and internally polls until finalized or a hard cap (~180 s, ceiling 240 s). On the cap it degrades to the non-blocking contract — it returns { still_running: true, call_id, … } and you continue with get_call. Use it when you want one call's result in one shot.
• Campaigns — 3–4 parallel, interchangeable, info-gathering-only calls (quotes, availability, inventory, policy comparison) — use place_call + get_call fan-out. Never use the blocking tool for campaigns. Never parallelize calls that book / buy / cancel / commit.

Caveats

• greeting does NOT set the agent's first spoken line. It is outbound-opener context only. The agent's actual opening line is a hardcoded recording disclosure on the server.
• 402 errors are PERMANENT gating, not retryable. trial_exhausted, plan_required, and reserved_number_required mean a human must act. Forward the returned action.url / sign_in_url verbatim (they are personalized and embed your proto-key). Do not retry.

Trial limits are read from the live quota object in the error envelope — they are never hardcoded in this package, so the displayed numbers always match the server.

Environment variables

More

• skill-companion/MCP-COMPANION.md — companion note for the ClawCall skill describing how the two share identity and workflow.
• ../skill/SKILL.md — the ClawCall agent skill (the procedural layer; this MCP server is the connectivity layer over the same REST API).