@chainrender/mcp

A Model Context Protocol (MCP) server for building ChainRender apps with AI. It exposes the ChainRender SDK as tools so an AI assistant that speaks MCP (in your editor or an MCP-capable chat) can compose, grade, compile, and validate fully on-chain, Grade-A, verifiable apps.

The principle: AI proposes, the SDK and chain verify

ChainRender is designed so AI output is checkable, not blind:

• Closed vocabulary — the AI can't express an invalid construct.
• grade() — the AI's trust claim is computed, not asserted (A/B/C + itemized breakdown).
• Deterministic compile() — you reproduce the exact hash the AI produced.
• Enforcer validation — an unsafe app is refused; the validate tool catches it locally first.

Tools

The server's instructions (sent when a client connects) are Start here for AI builders: the model does not know ChainRender from training data, where each fact is published, and the verify loop. Same text as the top of llms.txt.

Keep the context small: the full vocabulary is ~59 KB, most of it resolver kinds. Request every section except resolverKinds ({"include":["nodeTypes","a11y","behavior","enums","capabilities","kinds","limits"]}, ~5 KB) and look kinds up with chainrender_search_kinds → chainrender_describe_kind.

Install & run

npm install -g @chainrender/mcp
chainrender-mcp            # runs on stdio

chainrender_verify, chainrender_doctor and chainrender_review with chain read the published deploy records and the chain. Override any source with an environment variable:

Review a tree from the command line

For a tree an AI gave you, when that AI cannot run MCP (for example a chat model, or Google AI Studio):

npx -y @chainrender/mcp review app.json                  # human-readable report; exit 0 = pass, 1 = fail
npx -y @chainrender/mcp review app.json --chain sepolia  # also check every address against that chain
npx -y @chainrender/mcp review app.json --json           # the full result, as chainrender_review returns it

It is the same review as the chainrender_review tool and the visual builder's Review tab. Paste the printed report back to the AI and ask it to fix every item.

Configure in an MCP client

Add to your client's MCP server config (e.g. an editor's mcp.json):

{
  "mcpServers": {
    "chainrender": { "command": "chainrender-mcp" }
  }
}

The AI workflow

1. The AI calls chainrender_vocabulary (slim) to learn what it can compose, and chainrender_search_kinds / chainrender_describe_kind for the data sources it binds — encoding each arg exactly as the descriptor says.
2. It composes a Node tree from the vocabulary (using the compose.* builders' shapes).
3. It calls chainrender_check_node and fixes every issue — compile() silently drops unknown fields and invalid enum values.
4. It calls chainrender_grade to check the trust level — and fixes anything that lowered it.
5. It calls chainrender_verify for every fact it relies on (a kind, an export, a contract on the target chain, an address) and uses only what is Confirmed.
6. It calls chainrender_review (with chain) and does not report the app as done until pass is true and every honesty warning is fixed or explained.
7. The human deploys the reproducible, graded, validated result.

Every step after the AI composes is a verification, not trust. See the ChainRender docs: "Building with AI" and "Instructing an AI" — also as Markdown for agents at https://chainrender.org/docs/llms.txt.

Status

v0.9.0. Wraps the shipped @chainrender/client (^0.16.0) — the tools are thin wrappers over the SDK's vocabulary(), checkNode(), nodeJsonSchema(), grade(), compileComponent(), validateApp() and honestyFindings(); verify and doctor read chains.json, the Addresses page and the chain (src/chain.ts). Tests: npm test drives the built server over stdio against local mocks. No logic is duplicated; the SDK is the single source of truth.