Skip to main content
Glama
Yardhand
by escalay

Why

Every agent you run, whether Claude Code, Cursor, Codex, claude.ai or your own scripts, wants the same tools, and each one wants them configured again, with the keys pasted again. Yardhand sits in front of your MCP servers and REST APIs and gives every agent one address instead.

  • One address, every tool. Agents search for a tool by what they want to do, read its schema and run it. Hundreds of tools cost an agent a handful of tokens until it needs one.

  • Keys stay put. Each server's credentials are stored once, sealed with AES-GCM. Agents get keys of their own, each reaching only what you choose, read-only or not, and revocable on their own.

  • You decide what runs. Switch a tool, a group of tools or everything that deletes off. A third-party server that changes a tool's description waits for your review before agents see it.

  • Everything on the record. Every call, by which agent, with what result and how long it took, live in the app and optionally traced to Langfuse.

  • Packs. Hand an agent a curated kit for one kind of work, in your own wording, with a key that expires.

  • Yours to run. One container, SQLite, daily backups and restore in two clicks.

Related MCP server: Peta Core

Run it

mkdir yardhand && cd yardhand
curl -fsSLO https://github.com/escalay/yardhand/releases/latest/download/docker-compose.yml
docker compose up -d
docker compose logs gateway        # the setup code
open http://localhost:8080/app     # claim it, then a guided setup connects your first server and agent

No settings are needed to start. Self-hosting covers HTTPS, your own network, backups, updates and every setting.

On a new server it is one paste: give deploy/cloud-init.yaml as the user data when you create an Ubuntu or Debian server at any provider, and Yardhand is running when it boots.

Connect an agent

The app writes the setup for each agent, key included. For Claude Code it is one line:

claude mcp add --transport http --scope user gw https://your-yardhand.example/mcp \
  --header "Authorization: Bearer <key>"

Scripts use REST: GET /api/tools?q=… to find a tool, POST /api/calls to run it, or the typed client in packages/client. The API is described at /docs, and agents read /llms.txt.

How it fits together

flowchart LR
  agents["Agents"] -- "MCP · /mcp" --> yardhand
  scripts["Scripts"] -- "REST · /api" --> yardhand
  claude["claude.ai"] -- "OAuth" --> yardhand

  subgraph yardhand["Yardhand"]
    direction TB
    find["Search across every tool"]
    guard["Keys, switches and reviews"]
    record["A log of every call"]
    find ~~~ guard ~~~ record
  end

  yardhand --> mcp["MCP servers"]
  yardhand --> rest["REST APIs (OpenAPI)"]
  yardhand --> composio["Composio apps"]

docs/architecture.md has the pipeline every call goes through, with diagrams.

Telemetry

Once a day Yardhand sends one anonymous usage report: ranges and yes/no answers about how it is set up and used, with no names, addresses, queries or keys. The first goes 24 hours after the first start, and Settings shows the exact report before then. GATEWAY_TELEMETRY=off (or DO_NOT_TRACK=1) in .env switches it off. docs/telemetry.md lists every field, what the collector keeps and for how long.

Updates

Settings shows when a newer version is out, with its release notes. On a Docker Compose install, "Update now" has the updater beside the gateway verify the signed release, start it, and put the old version and your data back if it doesn't come up. The updater holds the Docker socket; GATEWAY_UPDATER=0 removes it, and GATEWAY_UPDATE_CHECK=off stops the daily check. docs/updates.md explains every step.

Develop

Node 26 and pnpm (corepack enable).

pnpm install
pnpm demo        # a gateway with fake servers and the app on http://localhost:4400/app
pnpm gate        # lint, types and tests
pnpm e2e         # browser tests

CONTRIBUTING.md explains how changes land and how releases are cut. Security issues go through SECURITY.md, never a public issue.

License

AGPL-3.0. Run it, change it and self-host it freely; if you offer a modified Yardhand to others as a service, share your changes under the same license. Illustrations are Open Doodles (CC0) and the typeface is Inter (OFL).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      A centralized management platform that aggregates multiple Model Context Protocol (MCP) servers into a single unified endpoint for AI agents. It provides a web interface for hot-swappable tool management, proxying of existing servers, and AI-powered generation of custom MCP plugins.
      1
      -
    • F
      license
      Not graded
      quality
      B
      maintenance
      A production-ready MCP gateway and control plane that provides credential vault, policy engine, audit logging, and managed runtime for routing tool calls between AI agents and downstream MCP servers.
      59
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to connect once to a local MCP endpoint and access tools from all configured MCP servers, add or remove connectors live, and generate sandboxed connectors for services that lack one. It centralizes multi-server access while keeping credentials isolated and calls logged without recording arguments or results.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      A self-hosted platform that aggregates upstream MCP servers, OpenAPI imports, and remote device connectors behind a single endpoint per profile, with per-agent scoped authorization, tool/prompt/resource catalogs, and approval-based temporary permissions. It also lets administrators distribute skills and connectors to paired devices and export client configurations.
      12,234,178 npm
      Apache 2.0