AMABIE / FIELD GUIDE

Start with
your own checks.

A practical guide to the current product. Built around executable evidence, with the version and trust boundaries stated explicitly.

Amabie 2.0.2 is available on npm.

The package is amabie-cli; the command is amabie. Clean installations are verified on Windows and Linux. Version 2.0.2 repairs baseline dependency handling: a check that cannot start is inconclusive. Active sessions are also recorded outside the project, so deleting local session files does not authorize resealing. Prompt-to-Evidence and authorization experiments remain opt-in; they are not general availability claims.

View the npm package · Review the verified runtime source
01 / SETUP

Inspect first. Then connect.

Install the pinned public release. You need Node.js 22 or newer, npm and Git.

npm install --global amabie-cli@2.0.2
amabie --version
amabie --help

Expected version: amabie 2.0.2. For a complete passing-change, regression and false-green walkthrough, use the packaged quickstart .

Inside a project you trust, inspect the discovered checks before sealing the setup. Project commands run with your local user permissions.

amabie setup --check
amabie setup --yes
amabie doctor --json

Review the checks and integration trust steps before setup. When upgrading, preserve your integration settings, trust stores and evidence; follow the quickstart’s upgrade steps instead of forcing a conflicting global executable. Full native macOS product-suite coverage remains unverified; its standalone build and version smoke test have been checked.

02 / WORKFLOWS

Three distinct operations.

Recommended: start with a reviewed failing test. Before the agent starts, the human writes a test for the required behaviour, confirms the expected failure, commits it and runs amabie setup --yes from a human terminal. Setup may report the intentional failure; the agent then implements the behaviour while keeping that test intact. For a separately registered objective criterion, the human binds the reviewed check before the agent starts: amabie bind test --requirement "<exact criterion>" --reseal. The intentional red baseline remains blocked until implementation.

If the agent writes a regression test, keep existing checks intact and add a new test file that the sealed command actually runs. Amabie requires failure on the starting implementation and success with the change. Worker-authored evidence carries an origin caveat. Rewriting existing checks, including adding assertions to an existing test file, may remain NOT PROVEN. Equivalent refactors may also remain unproven when checks pass on both sides.

Everyday completion

The agent registers each new request with amabie_task, preserving explicit criteria. Registration adds duties; it has no proof authority. The completion hook runs the sealed checks and reports remaining proof gaps.

Isolated work

amabie work opens a candidate for expert workflows. Objective requirements need sealed bindings. amabie finish verifies and promotes only when the required proof holds. Human acceptance does not replace objective proof.

Read state or run checks

amabie result --json and status read known state. doctor runs checks now. A focused doctor --check reports PARTIAL; it does not certify the complete task.

The attested baseline/candidate pipeline is a separate, stronger verification surface. It classifies repeated execution facts and retains an Evidence Bundle. A bound requirement is coverage inside this model, not a guarantee of semantic correctness.

03 / EXPERIMENTAL · CODING ONLY

Prompt-to-Evidence is opt-in and review-required.

A separate experimental workflow can use a local model to propose exact, source-linked requirement candidates from a coding prompt. Each candidate remains unclear and has no criteria until a person reviews it, resolves ambiguity, and binds it to executable evidence. The model cannot set a verdict.

With explicit authorization, a configured coding agent can receive a structured report of unresolved evidence for up to three correction attempts. Amabie reruns the complete sealed checks after each change. Scope changes, missing evidence, failed checks, or a reached limit remain unresolved. This experiment is not part of the Amabie 2.0 stable candidate.

The shared task-neutral core has a mock document verifier for contract tests only. It is not a supported document-agent integration; browser, research, data-analysis, and video verifiers are not available.

04 / INTEGRATIONS

The human trust step is part of the contract.

Claude Code: setup installs completion and pre-tool hooks. Unregistered requests are blocked on the normal hooked tool path; unbound objective criteria remain NOT PROVEN. Project-scope MCP tools require the human’s separate one-time approval in Claude’s /mcp menu.

Since 2.0.1, the request hook supplies a standalone CLI alternative while MCP approval is pending: amabie task "<intent>" --requirement "<criterion>" --session <id>. Omit requirements when none were stated. Recognized Claude calls cannot run setup, bind, accept or uninstall, or directly rewrite authority files. An open session requires human terminal confirmation for administrative changes. These guards do not create an OS sandbox.

OpenAI Codex CLI: setup writes the Stop hook, but Codex will not run it until you review and trust it through /hooks. A written hook alone gates nothing.

Other CLI agents: the integration is advisory. A machine-readable response does not imply the ability to block completion.

Native check discovery exists for Node, Python, Rust and Go. Other command-driven projects can declare an argv-based project contract. Execution evidence and runner observation are different claim levels; unknown runner internals are not silently treated as observed.

05 / SECURITY

LOCAL is the current boundary.

Authority records are signed and stored outside the project tree. A process with the same user identity can still replace the complete store or reseal a different plan and baseline. The everyday workflow gates completion; it does not put the agent’s native editor tools inside an OS sandbox.

HARDENED is unavailable in the reviewed build. Provider and broker code is implemented, but only a current measurement of the required independent authority boundary can activate that claim. There is no promise of universal fail-open prevention, exhaustive semantic verification or current token savings.

The public package is amabie-cli and the executable is amabie. Follow the release quickstart when upgrading; preserve existing integration settings, trust stores and evidence.

Reviewed trust documentation
06 / THIS WEBSITE

Your code never enters this demo.

The illustrative demo runs in the browser. It executes no project code, sends no code or evidence to a backend, and uses no account, form, analytics tracker or application cookie. Copying a snapshot uses the browser clipboard only after you click the copy button.

The site host processes technical request data to deliver and secure the pages. See the privacy notice for details and the legal notice for the operator.

07 / EXTENSION STATUS

Providers and extension points.

Implemented interface: an AI provider can attach a summary without changing classification. The only enabled provider shipped in the reviewed source is a no-op; this is not a bundled live AI explanation service.

Planned adapters: the architecture describes optional future security, performance and code-context adapters. Those extension plans are not listed here as working product integrations.