Keiro × LangGraph

Wire Keiro's remote MCP server into a LangGraph ReAct agent. Keiro gives your agent 4 web tools (cost credits per call) plus 8 free docs/utility tools — all read-only.

Full documentation — architecture, every tool & endpoint, params, credits, rate limits, auth, troubleshooting: see docs/keiro-langgraph.md.

Tool	Credits	Purpose
web_search	~1–2	structured web results
web_research	~3–5	deep multi-source research + synthesis
extract_url	~1–3	structured extraction from a URL (JSON schema)
answer	~5	direct Q&A, web search + AI synthesis
list_endpoints, get_endpoint, get_rate_limits, get_auth, generate_code, get_mcp_tools, suggest_schema, check_credits	free	Keiro API discovery

Endpoint / auth

• URL: https://kierolabs.space/mcp/api
• Transport: Streamable HTTP
• Header: Authorization: Bearer keiro_<key> — KEIRO_API_KEY env var

initialize runs without auth; every tools/list and tools/call needs the header.

Prerequisites

• Python 3.10+
• A Keiro API key (starts with keiro_) from the Keiro dashboard → API Keys
• An OpenAI API key (the ReAct agent's LLM) — or set OPENAI_API_KEY

One-command install + setup

git clone https://github.com/Keirolabs-API/LangGraph-Keiro && cd LangGraph-Keiro && ./setup.sh

This clones the repo, then setup.sh creates a .venv, installs langgraph langchain-mcp-adapters langchain-openai, writes a .env from .env.example (prompting for your Keiro and OpenAI keys), and runs a wiring self-check. No network calls in the check — it only asserts the config/auth header logic. Re-running ./setup.sh is idempotent and won't clobber existing keys.

Run a query

./run.sh "What changed in the EU AI Act this month?"

run.sh loads .env and invokes keiro-integration/keiro_langgraph.py. Costs Keiro credits (the four API tools deduct per call) + your LLM spend.

Use in your own graph

import asyncio
from langchain_openai import ChatOpenAI
from keiro_langgraph import build_keiro_agent  # add keiro-integration/ to sys.path

async def main():
    agent = await build_keiro_agent(ChatOpenAI(model="gpt-4o-mini"))
    r = await agent.ainvoke({"messages": [{"role": "user", "content": "latest news on X"}]})
    print(r["messages"][-1].content)

asyncio.run(main())

build_keiro_agent loads Keiro's MCP tools via langchain_mcp_adapters MultiServerMCPClient and hands them to LangGraph's create_react_agent.

Verify the wiring without network or deps

python keiro-integration/keiro_langgraph.py check

Troubleshooting

• KEIRO_API_KEY not set — run ./setup.sh, or export KEIRO_API_KEY=keiro_... and ensure run.sh can read .env.
• 401 / Unauthorized from Keiro — the key is wrong, expired, or missing the keiro_ prefix.
• tools/list empty — the Authorization: Bearer … header isn't reaching Keiro; confirm .env has a non-empty KEIRO_API_KEY.
• LLM errors — OPENAI_API_KEY missing/invalid, or the model name in keiro_langgraph.py (gpt-4o-mini by default) isn't available to your key.
• Credit deductions you didn't expect — only web_search, web_research, extract_url, and answer cost credits; the 8 discovery tools are free.

Files

• keiro-integration/keiro_langgraph.py — server config, client + agent builders, demo, self-check.
• keiro-integration/README.md — short integration-only README.
• setup.sh / run.sh / .env.example — install/run helpers.
• libs/, docs/, examples/ — the upstream LangGraph monorepo (vendored so the integration is self-contained).

License

See LICENSE.