Agent Doctor
An MCP server that audits ERC-8004 agent registrations, diagnosing issues and providing fix lists.
README
Agent Doctor
An AI-agent-facing service that audits ERC-8004 agent registrations and tells owners exactly how to fix them.
Why: of 170,000+ agents registered in the ERC-8004 registries, only 3–15%
expose BOTH a valid registration file AND a live declared endpoint
(arXiv:2606.26028). Agent Doctor diagnoses
that gap: it resolves the on-chain identity, fetches and validates the
registration file, probes every declared endpoint for liveness, verifies the
ENSIP-25 bidirectional ENS link, and returns a graded report (A–F) with a
concrete, ordered fix list.
What it checks
| Check | Meaning |
|---|---|
identity.resolve |
tokenURI(agentId) on the IdentityRegistry (0x8004A169FB4a3325136EB29fA0ceB6D2e539a432, same address on Ethereum, Base, Arbitrum, Optimism, Polygon, Linea) returns a non-empty agentURI |
registration.fetch |
the agentURI (https / ipfs / data:) serves valid JSON |
registration.structure / core_fields / services |
registration-v1 shape: type, name, description, image, well-formed services[] (lenient — soft fields warn, not fail) |
registration.registrations |
registrations[] actually references this chain + registry + agentId (eip155:{chainId}:{registry}) |
endpoint.* |
liveness probe per declared service — generic HTTP GET; A2A additionally looks for the agent card (endpoint itself, then /.well-known/agent-card.json); MCP is reachability-only (no handshake in MVP); ENS/DID/email/agentWallet are identifiers, skipped here |
ens.*.bidirectional |
ENSIP-25: the registration file references the ENS name and the name carries the text record agent-registration[<ERC-7930 registry>][<agentId>] — either side alone fails |
Grades: A all pass · B warnings only · C endpoint/ENS failures ·
D broken registration file · F identity unresolvable.
Quickstart
cd agent-doctor
python3 -m venv .venv
.venv/bin/pip install web3 fastapi uvicorn httpx mcp pytest
CLI
.venv/bin/python -m agent_doctor check --chain base --agent-id 42
.venv/bin/python -m agent_doctor check --chain base --agent-id 42 --ens myagent.eth
.venv/bin/python -m agent_doctor check --ens myagent.eth # best-effort discovery
.venv/bin/python -m agent_doctor check --chain base --agent-id 42 --json
Exit code 0 for grades A/B, 1 otherwise (CI-friendly).
HTTP API
.venv/bin/uvicorn agent_doctor.api:app --port 8004
curl 'http://127.0.0.1:8004/check?chain=base&agentId=42'
curl 'http://127.0.0.1:8004/check?ens=myagent.eth'
curl 'http://127.0.0.1:8004/healthz'
MCP (stdio)
.venv/bin/python -m agent_doctor.mcp_server
Exposes one tool, diagnose_agent(chain, agent_id, ens, timeout). Claude Code
registration example:
claude mcp add agent-doctor -- /path/to/agent-doctor/.venv/bin/python -m agent_doctor.mcp_server
Configuration
AGENT_DOCTOR_RPC_<CHAIN>(e.g.AGENT_DOCTOR_RPC_BASE) — override the default publicnode RPCs; ENS always resolves on Ethereum mainnet.AGENT_DOCTOR_IPFS_GATEWAY— defaulthttps://ipfs.io/ipfs/.AGENT_DOCTOR_HTTP_TIMEOUT— default 10 s.
Tests
.venv/bin/python -m pytest # unit tests (offline)
.venv/bin/python -m pytest --live # + live read-only test against the real Base registry
Fixtures include the ENSIP-25 worked example: registry
0x8004A169…a432 on eip155:1, agentId 42 →
agent-registration[0x000100000101148004a169fb4a3325136eb29fa0ceb6d2e539a432][42].
Register Agent Doctor itself (owner-only steps)
The repo ships its own dogfood identity in identity/:
registration.json (valid per our own validator), mint_identity.py
(unsigned tx builder — never touches keys), ens_records.py (exact ENSIP-25
record printer). To go live:
- Registration file is hosted — this repo is public, so the canonical
agentURI is
https://raw.githubusercontent.com/Maxwell012/agent-doctor/main/identity/registration.json. Theagentdoctor.example.comservice endpoints inside it are still placeholders — replace them once the HTTP API is deployed somewhere real. - Fund a wallet on Base (a few dollars of ETH for gas).
- Mint the identity — build the unsigned tx and sign it in your own wallet:
Function:.venv/bin/python identity/mint_identity.py --agent-uri https://raw.githubusercontent.com/Maxwell012/agent-doctor/main/identity/registration.json --from 0xYourAddressregister(string agentURI)on0x8004A169FB4a3325136EB29fA0ceB6D2e539a432(Base, chainId 8453). Read the mintedagentIdfrom theRegisteredevent, then updateregistrations[].agentIdinregistration.json(currently a0placeholder) and re-upload. - Buy the ENS name —
agent-doctor.eth(checked free on mainnet 2026-07-13); register it in the ENS app, set its ETH address record to the minting wallet. - Set the ENSIP-25 text record (bidirectional link):
Set the printed key to value.venv/bin/python identity/ens_records.py --ens <yourname>.eth --chain base --agent-id <agentId>1in the ENS app; also update the ENSservices[]entry inregistration.jsonif the name differs from the placeholder. Optionally set profile text records (avatar,description,url). - Optional: buy the $1 verification certificate at veral.tech for an extra trust signal.
- Self-check until grade A:
.venv/bin/python -m agent_doctor check --chain base --agent-id <agentId> --ens <yourname>.eth
Design notes / known limits (MVP)
- Read-only chain access; no keys, no signing, no state.
- The deployed registry proxy does not answer
totalSupply()/ enumeration (observed 2026-07-09), so ENS-only diagnosis without an agentId is best-effort and warns instead of guessing. - MCP probe is reachability-only; a real MCP handshake (initialize/tools-list) is a natural next step.
- Registration schema is validated leniently on purpose — the EIP-8004 registration-v1 field list is not a closed schema.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。