Skip to main content

The AI harness

Zeq is mathematics infrastructure. An AI system is one consumer of that infrastructure — a very natural one, but not what the framework is built around. The harness is the surface through which that consumer works: the tools an AI or coding agent needs to discover the framework, compute against it, and build on it. When someone spins up a machine, the harness is what their agent gets.

Everything here is a way to use the framework. None of it is the framework itself — the maths, the proofs, the clock, and the audit chain stand on their own, exactly as the rest of these docs describe.

The MCP surface​

The whole framework is reachable over MCP (Model Context Protocol), so an agent can call it the same way it calls any other tool. Point an MCP client at POST /api/mcp with a key and the core tools are there:

ToolWhat it does
zeq_spin_upprovision a machine and get a real compute key
zeq_computerun a verified computation (the seven-step pipeline)
zeq_contract_templates / _deploy / _firebuild and run contracts
zeq_publish_pagepublish a channel
zeq_verifyre-verify a ZeqProof
zeq_pulse · zeq_list_operators · zeq_field_statuspublic reads, no key

A free key is a single call away — see the quickstart.

Anti-guessing tools​

The most important thing a harness gives an agent is a way not to guess. An agent that has to invent operator IDs or endpoint names wastes calls and fabricates. So the framework exposes discovery tools publicly, no key required, on every node:

Operator lookup

Find the right operator by intent instead of guessing an ID — the registry is searchable, and an unknown ID is refused, never faked.

Operator specs

Ask for an operator's exact inputs, units, and equation before calling it, so the first call is the correct one.

Library search

Search the framework's own surface for the endpoint or capability that fits — discovery instead of trial and error.

Because the node bundle is byte-identical across the lattice, these tools answer the same way on every node — an agent learns the framework once and it holds everywhere.

The agent layer (optional)​

For consumers building whole agent systems on the framework, the harness adds an optional layer — each piece grounded on a machine, so an agent's work is auditable like everything else:

Workspace

State an agent can keep.

  • Files — read/write a versioned file workspace under a machine
  • Sandbox — run code in an isolated, quota-bounded sandbox
  • Plans — draft, lock, and track a plan of work
  • Skills — synthesise a reusable capability deterministically, and attach it

Populations

Many agents, governed by the clock.

  • Hyperagent engine — spawn a population, advance it generation by generation under a fitness gate, converge or retire it
  • Every step is a contract transition on the audit chain, with a ZeqProof

The full agent and harness endpoints are catalogued in the reference.

Pulse — the AI chat surface​

There's also a hosted front door for the AI consumer: Pulse, at ai.zeq.me. It's the conversational entry point — the gateway a person taps to talk to a framework-grounded model — and it's wired to the same seven-step protocol and proofs as everything else, so a chat that produces a number produces a verifiable one. The Pulse orb that appears across the surfaces is the shortcut into it. Pulse is a surface on the framework, the same as any app — the model talks, the machine grounds.

The line we hold​

The harness makes the framework easy for an AI to use. It does not make the framework about AI. A language model here is a caller like any other — it gets the same honest refusals, the same hypothesis stamps on generated equations, and the same proofs it can't talk its way past. The framework grounds the AI, never the other way around.