Agentplane is an agent-agnostic evidence layer for AI-written code changes. It keeps task intent, approved plans, verification evidence, traces, and Agent Change Records inspectable in Git.

Does Agentplane replace AI agent frameworks?

No. Agentplane works around agent frameworks by coordinating workflows, traces, context, artifacts, and operational state.

Does Agentplane run locally?

Yes. The default quickstart runs locally, writes repo-local artifacts, and does not require account creation.

What does Agentplane record?

Agentplane records task intent, approved plans, workflow runs, traces, verification evidence, review packets, and Agent Change Records.

What are Agentplane traces?

Agentplane traces are structured records of workflow steps, context loads, model calls, tool calls, checks, outputs, and artifacts from an agent run.

What is local context?

Local context is repo-specific operational knowledge that agents can inspect and reuse without depending on fragile chat history.

What is an Agent Change Record?

An Agent Change Record is a machine-readable record of AI-assisted engineering work, including task intent, changed files, verification evidence, and review status.

Agent-agnostic evidence layer for code review

Audit trails forAI-written codechanges.

When an agent changes code, reviewers need more than a diff. Agentplane keeps the task, approved plan, checks, traces, and Agent Change Record inside the repo.

Agent-agnostic. Local-first. No account required. MIT licensed.

Try the 90-second quickstartOpen GitHub

repo evidenceready for review

review readyAI-written change with evidence

Task intent, branch state, verification, and review packet stay in Git.

task20260523-AB12CDapproved scope + owner

branchtask/.../repo-evidence-heroisolated worktree

checksdocs:site:checkrequired before review

recordacr.jsonmachine-readable handoff

Task READMEintent, plan, findings

PR packetscope, checks, diffstat

Trace exportsteps, tools, checks

$ npm i -g agentplane
$ agentplane init
$ agentplane demo
$ human reviews evidence

51 stars64 releasesv0.6.4MITTypeScriptlocal-first CLI

A diff is not enough to merge agent work.

A diff shows what changed. It does not show why the agent changed it, what scope was approved, what checks ran, or whether the result is ready for review.

Agentplane turns AI-written code changes into repo-local evidence: task intent, approved plan, verification, traces, commits, and Agent Change Records.

From uncertain diff to reviewable change.

Give reviewers a durable trail before asking them to trust an agent-authored pull request.

Before Agentplane

The agent changed files, but the approved scope is buried in chat.
Checks ran somewhere, but the result is not tied to the task.
The reviewer sees a diff and has to infer intent.

After Agentplane

The task README records intent, plan, owner, findings, and rollback.
The ACR links changed files to verification evidence.
The PR carries a review packet instead of a loose agent summary.

See what an Agent Change Record contains

Try the review trail in a repo you control.

No account. No hosted dependency. Start locally, generate the evidence shape, then decide what to automate.

Run the local loop first

npm i -g agentplane
agentplane init
agentplane quickstart
agentplane demo
agentplane acr validate <task-id> --mode local

AGENTS.md
.agentplane/WORKFLOW.md
.agentplane/tasks/<task-id>/README.md
.agentplane/tasks/<task-id>/acr.json
.agentplane/traces/<run-id>.json

AGENTS.md

Repository policy gateway for coding agents.

.agentplane/WORKFLOW.md

Workflow and verification contract.

Task README

Task intent, lifecycle, plan, status, and handoff evidence.

acr.json

Machine-readable Agent Change Record for review and tooling.

Trace JSON

Inspectable run timeline and exportable trace data.

Recipes

Reusable workflow overlays for TDD, security review, and docs work.

Read the Agent Change Record guide

Agent-agnostic by design.

Agentplane does not replace your model, coding agent, editor, or Git workflow. It gives any repository-capable agent the same harness boundary and evidence trail.

Agent-agnosticLLM-agnosticCLI-firstGit-nativeLocal-firstOpenTelemetry-friendly

Agentplane is not another agent framework.

Agent frameworks define agent logic. Model providers run inference. Prompt tools shape instructions.

Agentplane coordinates the operational layer around coding-agent work: workflows, traces, context, recipes, verification, artifacts, and Git evidence.

Agent frameworksDefine agent logic and application behavior.

Model providersRun model inference.

Prompt toolsShape individual instructions.

AgentplaneCoordinates workflows, traces, context, artifacts, and operational state around agents.

Why harnesses matter

A model alone is not an agent. Real agents need workflows, tools, context, state, verification, and traces. Agentplane gives that surrounding harness a local, inspectable structure.

Read the harness engineering guide

ContextWhat the agent needs to know.

ToolsWhat the agent is allowed to do.

WorkflowHow the run progresses.

VerificationHow work is checked.

TraceWhat happened and why.