v0.1 · 10 MCPs · 4 clients · 39 tests green

Connect your AI to
real tools.

Agent_mcp is the MCP server registry shared by 4 AI CLIs (opencode / codex / cursor / kimi). One manifest, installed everywhere — GitHub, Filesystem, Playwright, Postgres, Sentry, Context7, Sequential-thinking … one install command writes opencode.json + codex config.toml + cursor mcp.json + kimi mcp add.

Get Started → View on GitHub See Architecture →

ONE COMMAND · FOUR CLIENTS

agent-mcp install github --client all

The truth source is registry/github/manifest.yaml. Adapters translate it into opencode's mcp object, codex's [mcp_servers.github], cursor's mcpServers.github, and kimi mcp add. Every write to a client config gets a timestamped backup first.

Read the handbook →

$ agent-mcp install github --client all

  opencode/github: writes ~/.config/opencode/opencode.json
                   backup: opencode.json.bak.20260516T...
  codex   /github: writes ~/.codex/config.toml
                   [mcp_servers.github]
  cursor  /github: writes ~/.cursor/mcp.json
                   mcpServers.github
  kimi    /github: kimi mcp add github -t stdio
                   -e GITHUB_TOKEN=$GITHUB_TOKEN -- ...

10 MCPS

GitHub · Filesystem · Playwright · Postgres · Sentry ...

github

Issues, PRs, CI failures, code search — full github.com loop.

filesystem

Allowlist-bounded file IO — clean for cross-tool batches.

context7

Up-to-date docs for fast-moving SDKs (React, Next, Supabase, Stripe...).

playwright

Real browser navigation, clicks, screenshots, DOM reads.

sequential-thinking

Structured multi-step reasoning for complex refactors & debugging.

git

Local git blame / log / diff — pairs with GitHub MCP.

postgres

Read schema, generate SQL, sanity-check migrations.

sentry

Production errors → code fix. Shortest exception→repair loop.

figma

Read design tokens & component structure — not just screenshots.

linear

Read tickets, generate plans, write back status.

Full MCP reference →

$ agent-mcp doctor

== schema ==
  ok github v0.1.0       (P0, ops)
  ok playwright v0.1.0   (P0, frontend)
  ...

== binaries ==
  ok    github       -> npx (/opt/homebrew/bin/npx)
  ok    git          -> uvx (/Users/lute/.local/bin/uvx)
  ...

== env ==
  WARN  github       -> GITHUB_TOKEN (not set)
  WARN  postgres     -> POSTGRES_CONNECTION_STRING (not set)
  WARN  sentry       -> SENTRY_AUTH_TOKEN (not set)

== client config consistency ==
  opencode: 6 managed by agent-mcp (...)
  codex   : 6 managed by agent-mcp (...)
  cursor  : 6 managed by agent-mcp (...)
  kimi    : 6 managed by agent-mcp (...)

ENV-VAR DRIVEN

Secrets go through env. Never into files.

requires.env: [GITHUB_TOKEN] in the manifest is the truth. Adapters write the env-var name (not the value) into client config; the runtime inherits the actual value from the shell at start. Never a plaintext token in any file — the schema validator detects literal ghp_* / sk-* / AIza* strings and rejects.

See the full architecture →

For people who hate configuring MCPs four times.

🧰

You use more than one CLI

opencode + cursor + kimi all need GitHub MCP? One command does all four at once.

🔐

You fear leaking tokens

Schema auto-detects ghp_ / sk- / AIza literals and refuses. Env-name only.

🩺

You want to know what's missing

agent-mcp doctor tells you: env not set, binary missing, anchor drift.

Install an MCP. Write to four places at once.

agent-mcp install github --client all. Backup, anchor, clean uninstall — guaranteed.

Start the install guide →

Connect your AI to real tools.