@agentskit/mcp

Profile: major-package

Tags: agentskit · typescript · ai-agents

Expose AgentsKit tools as an MCP server — use them from Claude Desktop, Cursor, Windsurf, or another host that supports the documented 2024-11-05 tools bridge.

Verified proof

• Package metadata and tests live under packages/mcp/.
• MCP package guide: https://www.agentskit.io/docs/agents/tools/mcp
• Stability map: docs/STABILITY.md

How this fits the ecosystem

@agentskit/mcp is the bridge from AgentsKit tools and agents into MCP-compatible coding agents.

• AgentsKit: expose tools, integrations, and registry agents through a standard MCP server.
• Registry: serve ready agents from registry.agentskit.io as callable MCP tools.
• Playbook: use playbook.agentskit.io for safe tool design, approvals, and agent handoff patterns.
• AKOS: run the same model with enterprise deployment, governance, and observability at akos.agentskit.io.

Docs: MCP package guide · agent handoff

npx -y @agentskit/[email protected] --tools fetch,search

Then point your MCP host at the command. Claude Desktop, Cursor, Codex, and OpenClaw all launch the same stdio server with host-specific configuration.

Claude Desktop or Cursor (mcpServers / .cursor/mcp.json):

{
  "mcpServers": {
    "agentskit": { "command": "npx", "args": ["-y", "@agentskit/[email protected]", "--tools", "fetch,search"] }
  }
}

Codex (~/.codex/config.toml):

[mcp_servers.agentskit]
command = "npx"
args = ["-y", "@agentskit/[email protected]", "--tools", "fetch,search"]

OpenClaw:

openclaw mcp add agentskit \
  --command npx \
  --arg -y \
  --arg @agentskit/[email protected] \
  --arg --tools \
  --arg fetch,search
openclaw mcp doctor agentskit --probe

Verify the server in each host with its MCP tool-list command before enabling additional tools. Pi does not have a native MCP host path in this recipe; use a Pi extension or package when a dedicated integration is available.

Flags

| Flag | Effect | |------|--------| | --tools <a,b> | which tools to expose (default fetch,search). Available: fetch, search, filesystem, shell, sqlite | | --fs-root <dir> | enable the filesystem tool, rooted at <dir> | | --sqlite <file> | enable the sqlite query tool against <file> | | --allow-shell | enable the shell tool (off by default — it runs commands) |

stdout is the MCP JSON-RPC channel; human output goes to stderr.

The bridge supports MCP protocol revision 2024-11-05 and the initialize/tools/list/tools/call lifecycle over stdio, in-memory, or an injected transport. Hosts own resources, prompts, sampling, tasks, HTTP/WebSocket, authentication, rate limiting, and persistence. Tool arguments are validated against their JSON Schema before execution; tools without a schema accept only an empty object. Invalid protocol envelopes and params are rejected, and tool failures are sanitized by default.

Programmatic

import { createAgentsKitMcpServer } from '@agentskit/mcp'
import { fetchUrl } from '@agentskit/tools'

createAgentsKitMcpServer({ tools: [fetchUrl()] }) // stdio by default

Pass transport to use a custom MCP transport (e.g. in-memory for tests).

Expose whole agents (agents as MCP tools)

Run a registry agent server-side and expose it as a single MCP tool — the host delegates a specialized job instead of orchestrating primitives:

npx @agentskit/[email protected] --agents legal-contract-reviewer,fintech-kyc-screener --provider openai

--provider covers the first-class adapters (openai, anthropic, gemini, ollama) plus the OpenAI-compatible set (deepseek, grok, groq, mistral, cohere, together, fireworks, openrouter, cerebras, kimi, huggingface, qwen, lmstudio, vllm, llamacpp) and any other OpenAI-compatible provider in the models.dev catalog. Key from --api-key or <PROVIDER>_API_KEY; --model / --base-url optional.

import { createAgentTool } from '@agentskit/mcp'
createAgentsKitMcpServer({ tools: [createAgentTool({ id, description, systemPrompt, adapter })] })

Typed Registry projections use fetchAgent + createTypedAgentTool; their input/output JSON Schemas are preserved. fetchAgentSkill intentionally rejects typed projections so they cannot be flattened into a text tool.

Quick start

import { createAgentsKitMcpServer } from '@agentskit/mcp'
import { createInMemoryTransportPair } from '@agentskit/tools/mcp'

const [, transport] = createInMemoryTransportPair()
const server = createAgentsKitMcpServer({ tools: [], transport })
await server.close()

Maturity and compatibility

• Stability: beta — see docs/STABILITY.md
• Node.js 20+ and TypeScript strict mode
• Published as @agentskit/mcp

Contributing

See CONTRIBUTING.md and the monorepo LICENSE.