Skip to content

Latest commit

 

History

History
80 lines (68 loc) · 6.89 KB

index.md

File metadata and controls

80 lines (68 loc) · 6.89 KB

interlock documentation

interlock checks every tool call your AI agent makes, inside the agent's own process, and allows, blocks, or rewrites it before it runs. These docs take you from a first guarded function to a production deployment.

New here? Do the getting started tutorial (about fifteen minutes), then read how interlock works. Running several machines and want one review queue and audit trail across them? Read the cloud plane; a single-machine install needs none of it.

How these docs are organised

Section What it's for Read it when
Getting started A guided first run, every step with real output You've never used interlock
Concepts How a decision is made: events, rules, verdicts, order, failures You want to understand what you're relying on
Security model What interlock protects against, what it assumes, where it stops Before you trust it with anything that matters
Guides (below) Step-by-step instructions for one task each You know what you want to do
Reference (below) Every function, parameter, detector, and CLI flag You need an exact name or default

Guides

Guide You'll learn to
Console Build and launch the local web UI over your running policy engine
Writing rules Block, rewrite, check results, scope rules to tools, and test your policy
Using detectors Pick from the 122 built-in detectors, make sure they see your tools, and start from a tested set
Declarative policies Write rules as JSON/YAML, lock the ones that must stay on, and accept project overlays safely
Approvals Hold a call for a human with ESCALATE, and write an approver
Coding agents Guard Claude Code, Codex and Cursor with one command
Credentials Let agents use secrets they never see
Databases and APIs Judge SQL by what it does, mask what comes back, keep personal data out of queries
Review queue Refuse now, let a human approve later, run the identical retry once
Integrations Connect @guard, MCP, LangChain, a model SDK loop, or a Node.js agent
MCP Guard MCP sessions and decide which servers and tools to trust
Scopes Give each tenant, agent, and sub-agent its own rules
Audit log Keep a tamper-evident record of every decision
Sandbox Run subprocesses under OS-level confinement built from your policy
Mesh Subscribe to community policy packs, verified from a pinned root and applied in monitor mode first; publish a commons; or report counts with Prio3 validity proofs over DAP
Production checklist Roll out safely: prove protection is on, monitor mode, refusals, logging, deployment
Cloud plane Run a hosted control plane for shared review, cross-host audit, and centrally compiled policy, without moving enforcement off the host
Rolling out a policy Compile, publish to a canary, watch il_policy_swaps_total, promote to 10% and 100%, and roll back without a redeploy
Releasing Cut a signed PyPI/npm release from a tag with trusted publishing and provenance attestations

What the mesh adds

Capability What it buys you Where
Signed commons Policy packs verified from a pinned root, with rollback and split-view defences Mesh
Monitor-first installs A new pack escalates on a canary and early rings, and blocks only where an operator grants it Mesh
Operator grants A pack the operator has not granted never blocks Mesh
Consent-gated reporting Nothing leaves a node until report enable, layer by layer, never retroactively Mesh
Secret-shared counts Aggregators see one additive share each; only their sum is a total, above a batch floor Mesh
Prio3 counts over DAP Validity proofs stop a node from lying about its counts; fixed, non-overlapping batches stop differencing; any DAP deployment can host the leader and helper Mesh
Transparency log and mirrors Every published entry is appendable and attributable; mirrors and gossip expose split views Mesh
Membership credentials A consensus service can require a credential before a report counts Mesh
Data licence and erasure Samples carry a warranty and a licence; an erasure request names what it removes Mesh
Offline root ceremony Root keys are held by distinct people, offline, with a published quorum Root ceremony

Reference

Page Covers
Python API Every public class and function, with signatures
Detector catalog All 122 detectors: what they catch, which calls they see, how to build them
Command line interlock doctor, report, hook, install-hook, policy, learn, trust, mcp, reviews, drift, mesh (including mesh dap task|submit|collect and serve --role dap-leader|dap-helper)
Sidecar The interlock-sidecar daemon for Node.js agents, and its wire protocol
Policy artifact Every field of the signed artifact a host enforces, canonical JSON, the epoch, the C ABI, and the parity and budget contract
Cloud runbook Key rotation, host freeze, chain fork, org deletion, restore and the version floor, manifest expiry, kill switch, evaluator rollback, and webhook triage
Node.js client The interlock-guard npm package
Native scope store The optional Rust backend for very large scope trees

More