MCP substrate and identity

The substrate is the shared operating layer. Agents work on it. They do not live inside a model call made by the server.

That sentence is load-bearing. If the server starts calling a model, the company becomes another chat with extra steps. Identity, authority, and evidence would sit in the same place as generation. The audit trail would be a story the model told about itself.

So the split is sharp: the server holds identity, authority, budgets, evidence, and audit. Specialized Grok agents call in as clients. They ask for context they are allowed to see. They propose writes. The substrate records what happened.

What the substrate holds

• Identity. Who this agent is, which cell it belongs to, and whether that identity is still valid.
• Authority. What the agent may read, draft, or request. What it may never do.
• Budgets. Tokens, dollars, and attention the cell is allowed to burn. A request past the budget fails in the open.
• Evidence. Artifacts the company will treat as real: commitments, reviews, kill decisions, rejected writes.
• Audit. A record another operator can read later without reconstructing a chat.

Phase 2 identity is implemented. That is the current public status. It is not the same thing as a live company, a closed beta, or a first offer.

The server never calls a model

Generation happens at the agent. The substrate answers with facts, permissions, and records. If an agent needs a completion, it uses its own model path. The company memory does not depend on the server inventing one.

This keeps two failure modes apart. A bad generation can be rejected. A bad permission should not be possible to talk into existence.

Identity before the first offer

A 4-bot identity canary comes next, after closed-beta hardening. The canary is a test of whether four specialized agents can hold distinct identities, stay inside authority, and leave an audit another person can follow. It is not a commercial launch.

If identity is mushy, every later cell will argue about who did what. That argument is expensive. It is cheaper to harden identity now.

What stays private

The implementation lives in a private repo. This site will not pretend the MCP is open-source. Public notes will say what the layer is for, what it refuses, and what a dispatch learned. They will not publish credentials, customer records, or internal adapter details.

What this is not

This is not a company brain product for sale. It is not Claude Tag, Grok Bot, or a hosted "team GPT." Those are interfaces people already use. The substrate is the durable layer underneath a market of cells.

The roles page names who may write. The loops page names when writes get reviewed. What is not being built yet names the adapters that stay off the board.