tctc-mcp

An MCP server exposing ERC-7303 (Token-Controlled Token Circulation) roles to AI agents: agents check their own on-chain permissions, and human principals grant/revoke them by minting/burning control tokens — no permission server required.

Status: v0.3 — adds timed roles with gasless auto-expiry (below), on top of v0.2's IERC7303 auto-discovery; published on npm (tctc-mcp), unit-tested and verified end-to-end against the Sepolia demo deployment (grant → check → revoke → check, and grant-for-75s → auto-expiry, through a real MCP client).

Demo (60 seconds)

A human grants an AI agent a minting permission; the agent verifies it on-chain and mints an NFT. The human burns the role token — and the agent instantly loses the capability. Live on Sepolia, no permission server involved.

Quick start

The package is published on npm, so no clone or build is needed — npx fetches and runs it directly:

# 1. Get a config. The secret-free Sepolia demo config needs no API keys:
curl -fsSLO https://raw.githubusercontent.com/kofujimura/tctc-mcp/main/examples/config.sepolia.agent.json

# 2. Register with your MCP client, e.g. Claude Code
#    (read-only mode: only query tools are registered)
claude mcp add tctc -- npx -y tctc-mcp --config "$PWD/config.sepolia.agent.json"

# Admin mode (principal side): grant_role / revoke_role also registered.
# Provide the issuer key ONLY via the environment:
claude mcp add tctc-admin --env TCTC_ADMIN_PRIVATE_KEY=0x... \
  -- npx -y tctc-mcp --config "$PWD/config.sepolia.json"

Or in a project-scoped .mcp.json:

{ "mcpServers": { "tctc": { "command": "npx",
    "args": ["-y", "tctc-mcp", "--config", "examples/config.sepolia.agent.json"] } } }

A fuller registration example is in examples/claude.mcp.json. The admin private key is only ever read from the TCTC_ADMIN_PRIVATE_KEY environment variable; configs containing anything that looks like a private key are rejected at startup.

Using tctc-mcp from your own app

The only supported entry point is the tctc-mcp bin — spawn it via npx --no-install tctc-mcp (or node_modules/.bin/tctc-mcp). Never reference the package's dist/ files directly: they are internal, their layout is unversioned, and the package's exports field refuses deep imports. A minimal MCP-SDK client example (the pattern for a web backend) is in examples/client-stdio.mjs; the release pipeline verifies the packed bin end-to-end (scripts/verify-pack.mjs).

Web application starter

Want to see tctc-mcp protecting a real API route?

tctc-openai-starter is a standalone Next.js starter that verifies wallet ownership with Sign-In with Ethereum, checks the configured role through tctc-mcp, and calls OpenAI only when the on-chain gate grants access.

Use it as a GitHub template.

Token-gating any MCP server: tctc-gate

The reverse direction lives in this repo too: tctc-gate (npm) is a transparent stdio proxy that puts an ERC-7303 role check in front of any existing MCP server, unmodified — deny comes back with a grant URL that opens the dashboard pre-filled. tctc-mcp lets agents ask about roles; tctc-gate enforces them at the boundary:

npx -y tctc-gate --config gate.json -- npx -y some-mcp-server …

Tools

* registered only when the config has an identity section.

Subjects can be given as a raw address, as an ERC-8004 agentId (resolved to its ERC-6551 Token Bound Account, the recommended binding target), or omitted to use the config's self.

IERC7303 auto-discovery (v0.2)

ERC-7303 now defines an introspection interface (ethereum/ERCs#1872, merged 2026-07-11): compliant contracts expose hasRole, control-token getters, configuration events, and ERC-165 detection (interfaceId 0x4ee69337). tctc-mcp uses it two ways:

• target roles — a role config names only the target contract; the server reads which control tokens gate the role from the contract itself, and the verdict is the target's own hasRole() answer:

  json

  Copy

  "roles": { "MINTER_ROLE": {
      "target": { "address": "0x4C0a78803D47154B9C6F42EC4AEbab2D1C94c97D" } } }
  

• discover_roles — introspect any address at run time, with no role configuration at all. Non-compliant contracts report supportsIERC7303: false; static controlTokens configs remain the fallback for pre-IERC7303 deployments.

Working example: examples/config.sepolia.discovery.json (secret-free, public RPC), verified live by scripts/e2e-discovery.mjs.

Timed roles: gasless auto-expiry (v0.3)

Delegation to an agent is usually short-term — "mint for one hour", "act for the duration of this task". With an expiring control token (ExpiringControlTokens), balanceOf() returns 0 once the holder's expiry passes, so the role revokes by itself, with no transaction — even if the principal forgets, goes offline, or loses keys. The ERC-7303 target contract needs no changes at all (the Sepolia expiry demo target is a byte-for-byte copy of TCTCDemoToken).

• Granting: a role whose grant template has $expiresAt requires an expiry — grant_role with expiresInSeconds: 3600 is "grant MINTER_ROLE for one hour":

  json

  Copy

  "admin": { "grant": { "function": "mint(address,uint256,uint64)",
                        "args": ["$subject", "$typeId", "$expiresAt"] } }
  

• Checking: check_role evidence reports expiresAt (unix seconds) when the control token exposes it, so an agent can self-report "this permission expires in 5 minutes".

• Kill switch unchanged: expiry is a fail-safe, not a replacement — revoke_role (issuer burn) still revokes immediately within the validity window.

Working example: the TIMED_MINTER_ROLE in examples/config.sepolia.json, verified live by scripts/e2e-expiry.mjs (grant for 75 s → watch it expire with no further transaction).

Human dashboard

Live: https://tctc-mcp.vercel.app/ — while tctc-mcp exposes ERC-7303 roles to AI agents, the TCTC Dashboard exposes the same on-chain state to the humans who manage them: inspect any IERC7303 target, watch hasRole verdicts and per-control-token balanceOf evidence live, grant/revoke as the issuer (with timed grants and countdowns for expiring control tokens), and deploy new control-token collections straight from a browser wallet. Two clients of one source of truth: the chain. See dashboard/README.md for details and try-it links.

Not just for AI-agent delegation: any ERC-7303 contract works, so the dashboard doubles as a general-purpose on-chain permission manager — issue certificate collections, grant and revoke roles, and audit who holds what, all from a browser wallet.

Documents