Documentation
Everything you need to run Atested
From install to production. The proxy, the classifier, the policy, the chain, and the dashboard.
Getting started
Installation guide
Clone the repo, start the proxy, point your agent. Five minutes to a running governance layer with a signed chain.
Read the guide →
Provider setup
Configure Atested for Anthropic, OpenAI, Gemini, or LiteLLM. One environment variable per provider.
See provider setup →
For developers
The technical path. Install, integration, policy, CLI commands, and verification — concrete, with commands and file paths.
Developer guide →
How it works
Architecture
How Atested fits between your agent and the model provider. The interception point, what gets inspected, what passes through.
How the proxy works →
The classifier
How Atested classifies operations by evidence. Action types, target classes, confidence tiers.
How the classifier works →
Policy rules
Declarative rules, first-match evaluation, and the governance logic that ships with Atested.
Policy rules reference →
The chain
Append-only, hash-chained, Ed25519 signed audit records. How the chain works and how to verify it.
How the chain works →
Integrity monitor
Chain file checks, proxy source hashing, policy rules hash verification, and the events that document runtime integrity.
Integrity reference →
The decision
Every classification meets a policy rule and produces an ALLOW or DENY. The decision is deterministic — same input, same rule, same outcome. Both decisions are signed and recorded with the evidence that drove them.
How decisions work →
Using Atested
The dashboard
Activity, approvals, reports, alerts, health, configuration, and audit in one place.
Using the dashboard →
Approvals
When Atested can't classify an operation, you decide. Approve once, Atested remembers. If it changes, it resets.
Managing approvals →
Agent governance
How Atested governs OpenClaw, Claude Code, Cursor, and other AI agents. Auto-onboarding and agent-specific guidance.
Agent governance guide →
Configuration
Policy rules, base directories, discovered tools, signing keys, and proxy settings.
Configuration reference →
Reports and support
Predefined reports, JSON export, Telemetry Transparency, Trouble button context capture, and Trouble History.
Reports reference →
Restore from backup
Primary restore procedure, required gov_runtime contents, and restore verification for multi-machine installs.
Restore guide →
Trust and transparency
Deep dives
Design rationale
Why an HTTP proxy, why evidence inference, why confidence tiers, why the chain. The choices and their tradeoffs.
Read the rationale →
Why governance
What goes wrong without governance, what governance adds, and why deterministic verification beats judgment.
Read the argument →
How Atested got built
Atested is built with AI. Not assisted by AI. Built. This is how the discipline works.
Read the story →
Quick reference
Supported providers
Anthropic, OpenAI, Gemini, LiteLLM
View reference →
Environment variables
All config options and their defaults
View reference →
Chain record format
JSON schema and field reference
View reference →
Confidence tiers
Tier 1–4 classification reference
View reference →
Telemetry
What we collect, how to opt out
View reference →
GitHub repository
Source code, issues, releases
View repository →
Start attesting your AI operations
Stop worrying and start knowing in less than five minutes.