Open-source ToolYour MCP

Remote MCP gateway for AI agents: plan → run playbook/solve → verify until pass.

MIT licensed. Hosted API/SaaS backends are optional — offline evals and local file control-plane run without them.

Product docs: https://www.toolyour.com/developers/mcp
Official registry: com.toolyour/mcp
npm client (separate): @toolyour/sdk (MIT)

Quick start

cp .env.example .env
npm install
npm run build:registry   # from monorepo docs/ when regenerating
npm run build
npm run dev

Endpoints: http://localhost:3090/mcp (legacy SSE for compatible hosts) · Streamable HTTP: /mcp or /mcp/http. All Streamable HTTP endpoints are stateless: clients do not need session affinity, and a stale Mcp-Session-Id is ignored. The OAuth-protected /mcp/chatgpt endpoint validates its bearer token on every request. GET /mcp remains the legacy SSE endpoint and keeps its session behavior.

Health: GET /health/mcp · Ready: GET /health/mcp/ready · Metrics: GET /health/mcp/metrics (Prometheus text)

Related MCP server: mcp-graph-loop

Auth

X-Api-Key: ty_...

Same key as ToolYour REST. Free tier: 500 credits/month (tools cost 1–10 credits). MCP exposes API-backed catalog tools only (hasApi).

Host contract

ToolYour does not replace Cursor/Claude. The host keeps editor, git, and terminal. Agents should:

1. plan_task (free)

2. run_playbook or solve_task

3. Apply only rank-1 loop.nextActions in the workspace

4. verify_task with prior result as baseline until loop.gate is pass (or stop)

Skill-loop responses also include an additive execution envelope (runId, intentId, intentType, idempotencyKey, and projectScope). Preserve it when handing a result to another MCP tool. Async jobs reuse execution.runId as the get_run polling id.

Internally, plan_task, solve_task, run_playbook, run_workflow, and verify_task use one executeIntent lifecycle boundary for correlation and the execution envelope. Credits settle on each gateway tool invoke (billing.settledBy = gateway_per_tool) — MCP does not double-bill. Free recall_context recalls purpose-typed memory before rebuilds.

Saved Playbooks (private beta)

Saved Playbooks turns a repeatable ToolYour workflow into a named, versioned recipe that an MCP-capable agent can run again. V1 deliberately points only to existing ToolYour skills or workflows; it does not store arbitrary shell commands, control another agent, or bypass host approval. Each account gets owner-scoped playbooks and durable run receipts.

Enable this only on a private/default MCP deployment with PLAYBOOKS_BETA=true. The beta adds create_saved_playbook, list_saved_playbooks, get_saved_playbook, update_saved_playbook, run_saved_playbook, and get_saved_playbook_run. It is intentionally excluded from /mcp/chatgpt so the reviewed public tool surface stays separate. The public surface includes the authenticated get_profile identity tool.

See Saved Playbooks beta design for the data model, safe execution boundary, and rollout guidance.

Offline evals (no API key)

npm run ci                 # build + lint + unit/contract/integration + evals
npm run eval:golden        # routing + jobReport fields + verify gate
npm run eval:playbooks     # skill → workflow → synthesizer matrix + fixture coverage
npm run golden:tier1       # host work-package shapes

Fixtures: tests/eval/synth-fixtures.json · Goals: tests/eval/goals.jsonl

Observability (open protocols)

No proprietary APM required.

What stays hosted (optional)

• Live tool execution via gateway + API key

• Feature Memory / verification profiles (SaaS Mongo)

• Redis for multi-replica async get_run

Local defaults use file-backed control-plane jobs.

License

MIT — see LICENSE.