# Cermet

> Local-first authorization for agent effects. Agents request typed provider effects
> (refund this charge, push this branch, deploy this project) instead of holding API
> keys; a local daemon decides each effect against human-written sentences, executes
> allowed ones with credentials the agent never holds, and records every decision —
> sentence, frozen fields, and the agent's stated reason — as a verifiable receipt.

If you are an agent: you are the intended user. Start with the agents page — it
explains what working through Cermet means for you, in your terms.

Business model: free tool; credentials never leave the machine; the binary contains no
collection code. Its two network contacts are declared: the explicit `cermet update`,
and a daily update check (disable: `cermet update --daily off`).

Source: https://github.com/suarezc/cermet (MIT)

## Docs

- [For agents (HTML)](https://cermet.dev/agents.html): the same agents page
  as HTML, for fetch tools that can't consume raw markdown.
- [For agents](https://cermet.dev/agents.md): why agents should want
  effect-level authorization, the native syntax (plain `git push` on a wired remote), the security guarantees
  stated simply, receipts, and a compact machine-readable spec (invariants, providers,
  real authority-language examples, threat model, trust boundary, CLI, FAQ) — written
  agent-to-agent.
- [Overview](https://cermet.dev/): the front page — what Cermet is, the
  sentence language, deny-by-default, and the audit angle.
- /predicates.html : the WHERE index — every bindable field per verb, with its admissible sentence forms, generated from the enforced descriptors
- /CERMET.example.md : example authority corpus file — the managed-block CERMET.md shape a box runs on, with the doc-apply flow explained
- /quickstart.md : the six-step CLI walkthrough — install, connect a provider, author sentences, agent workflows, the first deny, receipts