Hermes Plant MCP Server
[Glama score]
[CI]
Runnable MCP server and integration examples for Hermes Plant, the Agent Commerce Assurance layer that lets AI agents preflight, approve, and prove consequential actions. Start with a one-cent action-safety check, route only high-risk actions into signed review evidence, and pay per call over x402.
What's here: a runnable stdio MCP bridge for registry crawlers and local clients, plus drop-in examples in curl, TypeScript, Python, CrewAI, LangChain, and MCP client configs for Claude Desktop / Cline / Cursor.
Review one workflow before launch
An agent receives approval for one action, its arguments change, and a retry after
a timeout may execute it twice. What binds approval to the final action, prevents
duplicates, and proves whether execution completed?
The sample findings memo
walks through a synthetic example of these gaps, the available evidence, and a
ranked fix plan. It is a demonstration, not a client result.
For technical founders and automation agencies with a working or near-launch
agent, the $300 workflow review
covers one workflow, a workflow map, ranked findings, and a practical fix plan.
The three-hour cap includes calls, preparation, the memo, and one clarification
round. We agree scope and timing in writing before billing. Implementation and
production access require a separate agreement.
Install the Action Safety skill
Install the public, fail-closed Action Safety skill for Codex, Claude Code,
Cursor, and other Agent Skills-compatible clients:
The skill calls the one-cent quick gate before an exact consequential action,
binds the result to that unchanged action, and escalates high or critical risk
only when the additional $0.25 is authorized. Review triage is never treated as
human approval and the skill never expands the caller's permissions.
Source: skills/protect-agent-action · Install guide and trust boundary
Use the hosted MCP connector
For the complete production tool surface, use the canonical Hermes Plant — Agent Commerce Assurance connector on Glama. It exposes Action Safety, Spend Assurance, signed evidence, payment-policy checks, and the component x402 tools through the live Streamable HTTP endpoint.
This repository is the portable, no-secret discovery bridge. It lets registry crawlers and local clients inspect the live catalog and payment contracts without signing a wallet message or spending funds. The hosted connector is the buyer-facing path for invoking the production tools.
Glama registry
[Hermes Plant MCP Server Glama card]
This repo is arranged for Glama to build and inspect the MCP server without secrets or funded wallets:
glama.json declares the GitHub maintainer.
- The root
Dockerfile starts the stdio MCP server in mcp-server/.
npm run smoke:mcp lists all local MCP tools and fetches the live x402 manifest.
- GitHub Actions validates the MCP server, Docker image, shell examples, and no-spend TypeScript/Python client contracts with their real dependencies.
A Glama release is still an account-side action, not a GitHub release. After claiming the server in Glama, use the Dockerfile admin page to deploy the build, wait for the build test to pass, then publish a Glama release version. That release unlocks Glama's Server Coherence and Tool Definition Quality scoring.
Complementary MCP servers
Hermes Plant governs consequential actions and payments; these independent servers cover adjacent layers of an agent stack:
These links describe functional complementarity, not partnerships or endorsements.
Start with an assurance workflow
Use the free API key for up to 250 calls/month, the $29/month Agent API Pass for regular volume, or per-call USDC on Base. Public proof and machine-readable contracts are available at hermesplant.com/proof and hermesplant.com/openapi.json.
Component catalog
Nineteen hosted x402 endpoints remain available as composable utilities:
Agents discover endpoints through OpenAPI, llms.txt, the x402 manifest, the API catalog, MCP metadata, or agent skills. The workflow links above identify this GitHub repository as their source; actual Glama placements use separate registry tags. Source tags are observational and do not establish buyer identity or a completed purchase.
How an x402 call works
- Client makes an HTTP request.
- Server replies with
402 Payment Required plus a PAYMENT-REQUIRED header carrying the price, network, asset, recipient, and timeout.
- Client parses the challenge, signs a USDC transfer authorization locally, and replays the request with a
PAYMENT-SIGNATURE header.
- Server verifies the signature via the facilitator, runs the work, settles the payment on-chain, and returns
200 OK with the result plus a PAYMENT-RESPONSE header.
The full spec lives at github.com/x402-foundation/x402.
Discovery surfaces
These let any agent or human self-onboard without help:
MCP tools in this repo
The runnable stdio server exposes five read-only discovery tools:
These local tools do not sign wallet messages, approve transactions, call paid endpoints, or spend USDC. They expose the discovery layer that clients need before invoking paid hosted tools.
Quickstart
Pick the runtime that matches your stack:
Run the MCP server locally
MCP server registry build
This repo includes a root glama.json and root Dockerfile for MCP registry crawlers:
The container starts the stdio MCP bridge in mcp-server/, which exposes Hermes Plant discovery tools and hosted MCP metadata without requiring API keys.
Example source labels
The curl, TypeScript, Python, CrewAI, and LangChain CashflowLens examples send
x-hermes-channel: github-mcp-repo, x-hermes-medium: repository, and
x-hermes-campaign: first-workflow. x-hermes-content identifies the runtime.
The paid clients pass these headers into the x402 request so the SDK retains them
when retrying after a 402 challenge. Curl only inspects the unpaid challenge.
If you publish an example in another placement, replace these labels with that
placement's source and campaign. Labels are caller-declared observations; they
do not verify buyer identity, revenue, or an unaffiliated customer. Owner QA must
also send x-hermes-probe: 1 so it is excluded from acquisition reporting.
Wallet setup
The guarded paid-call examples require:
- A funded wallet on Base mainnet with chain id
8453.
- Enough USDC for the selected endpoint and a small amount of ETH for gas.
EVM_PRIVATE_KEY plus the explicit opt-in HERMES_ALLOW_PAYMENT=1.
The included CashflowLens clients cap payment requirements at $0.20 USDC and register only Base mainnet. Leave the opt-in unset for imports and tests. Never commit a private key. These production examples do not claim testnet support; use an x402 test resource server for testnet integration work.
Status
These examples target the production deployment at hermesplant.com. Endpoint signatures may evolve, so cross-reference openapi.json and the /.well-known/x402 manifest before going to production.
Contributing
Issues and PRs welcome. New runtime? New framework? Open a PR with one working example and a tight README.
License
MIT - see LICENSE.