MCP server — @zeq/mcp
The MCP server exposes the Zeq engine over the Model Context Protocol, so any MCP-compatible client — Claude Desktop, Claude Code, Cursor, Zed, and others — connects and becomes a Zeq agent. It spins up its own machine, gets a real key, runs verifiable physics, and builds and fires contracts and apps — all through tool calls, all landing on an audit chain.
Connect over stdio (default) or HTTP at POST /api/mcp.
The tools
| Tool | What it does |
|---|---|
zeq_spin_up | provision your own machine and get a real key |
zeq_compute | run a verified computation; returns a ZeqProof + a verify object |
zeq_contract_templates / generate / deploy / list / get / fire | build and run contracts |
zeq_publish_page | publish a channel app |
zeq_observer | the proof-of-computation feed |
zeq_verify · zeq_pulse · zeq_list_operators · zeq_lattice · zeq_shift · zeq_field_status | verify, clock, discover, and more |
Self-onboard in two steps
Every key belongs to an account, so onboarding is authenticated — but it's still two steps with no signup:
POST /api/demo-key/mint— free, no signup — to get a demo key.- Call
zeq_spin_upwith that demo key in theAuthorizationheader — it provisions your own machine and returns a real, activezeq_ak_key (free tier) plus your recovery equation, shown once.
Put that key in your connector and every other tool runs on your machine. From there, zeq_compute
returns a signed ZeqProof and a verify object you can hand straight to zeq_verify for a guaranteed
round-trip.
Rules and instruments
The MCP server is the instruments side of the loop — the callable tools. Pair it with the kernel skill, which is the rules side — the system prompt that pins the clock, mandates KO42, and wires the seven-step protocol. Give an agent both and it has the contract and the tools to honour it. The framing — that AI is a consumer of the framework, grounded by its proofs — is in the AI harness.
WebMCP — the same node, from inside the browser
POST /api/mcp is the server-side surface, for programmatic clients. Every page the node
serves also exposes tools to the in-browser agent through WebMCP — the W3C
navigator.modelContext browser standard (Community Group draft, Chrome 146+). This is a
distinct mechanism from the HTTP MCP above, not a transport for it: the page registers its own
tools, and a browser-resident agent calls them directly, in the visitor's own session.
The node ships this layer at /zeq-webmcp.js (feature-detected, injected on every page). It
registers five tools:
| WebMCP tool | What it does |
|---|---|
zeq_pulse | live Zeqond tick + commons-pool balance |
zeq_wallet | the signed-in ZID's compute-credit balance + conversion rate |
zeq_convert_credits | credits → ZEQ Coin — human-confirmed via requestUserInteraction (it moves value) |
zeq_api_catalog | discover the public endpoint surface |
zeq_call | call any public /api/* route (GET free; writes confirmed; admin/zsc/mcp blocked) |
WebMCP tools run against the visitor's same-origin session cookie and never forward keys.
Feature-detect with "modelContext" in navigator; the @mcp-b/global polyfill fills older
browsers. GET /api/mcp advertises the layer in its webmcp field.
The economy an agent meets
Beyond the autonomous per-Zeqond mint, metered compute is billed in compute-hours: 1 hour =
1 compute credit = 1 ZEQ Coin (opt-in conversion from the commons pool, supply-neutral). Per-seat
annual subscriptions buy those hours (Starter 10h · Builder 40h · Advanced 120h · Architect 400h),
and an annual seat is a lifetime for its version number. AI is strictly BYOK — the platform
never funds LLM inference out of compute hours — and credit balances are federated to the
credit-authority node. GET /api/mcp reports all of this under surface.economy.
Read next
- Kernel skill — the rules to pair with these tools.
- The AI harness — how AI consumes the framework.
- Proofs — what
zeq_verifychecks.