AgentSmith Get started

FREE AND OPEN SOURCE · LOCAL FIRST

The open-source harness for coding agents.

From vibe coding to agentic engineering. Proof before done.

Give Claude Code or Codex your project rules, a definition of done, and checks that produce evidence. Keep changes reviewable, carry decisions into the next session, and turn observed failures into validated improvements to your harness.

Try one verified task Get started for free

AgentSmith is MIT licensed. Your coding-agent subscription or model usage costs are separate.

How AgentSmith works

  1. Your taskScope + definition of done
  2. Claude Code or Codex + AgentSmithProject rules · configured checks · handoffs
  3. Inspect the resultChanges + retained evidence

  1. Understand and bound

    • Read existing code and callers.
    • Define scope and completion checks.
    • Make small, reviewable changes.
    • Choose native permissions explicitly.
    • Enable optional secret-scanning hooks.

  2. Verify and keep proof

    • Reproduce the failure before fixing.
    • Run all configured project checks.
    • Exercise the user-visible path.
    • Record check results in a receipt.
    • Inspect what the evidence covers.

  3. Continue and improve

    • Save decisions and a clear handoff.
    • Resume from project-owned records.
    • Capture observed failures.
    • Propose a rule, hook, or test.
    • Validate the change before adopting it.

Instructions guide behavior. Configured checks enforce specific boundaries.

02A CONCRETE EXAMPLE

The failing check stays visible.

In the public First Verified Loop, a readiness test catches a false “ready” result when one required check is missing. The corrected task passes syntax and tests, then a real command reports NOT READY for the incomplete input.

  1. Failing test
  2. bounded fix
  3. configured verification
  4. real invocation
  5. receipt
  6. handoff
Run the disposable demo Read the full evidence bundle
  1. 01Red testFailure before the fix
  2. 02Green verificationThree configured phases pass
  3. 03Real pathThe command reports NOT READY
  4. 04Receipt + handoffRecorded result and resumable state

03THE WORK LOOP

A path from task to evidence.

AgentSmith supplies the structure; your project supplies the checks. The completion claim is tied to something a person can inspect. When a failure is observed, it becomes a proposed change, and you verify that change before you adopt it.

  1. Project rules and a profile define boundaries and done.

  2. The agent works on one bounded task.

  3. Automated checks and a real invocation test the result.

  4. Evidence supports the completion claim.

  5. A handoff lets another session resume safely.

  6. An observed failure becomes a rule, hook, or test you validate.

04GUIDED SETUP

Set up AgentSmith.

Choose one route. Both use the same guided setup and produce the same project files.

ROUTE 1

Download and run guided setup

  1. Download the standalone build for macOS, Windows, or Linux.
  2. Open a terminal and run agentsmith.
  3. Answer the guided questions. AgentSmith shows the exact plan before it writes anything.

The standalone app includes its runtime; Python is not required. The macOS and Windows builds are currently unsigned, so your operating system may ask you to confirm before running them.

Open the latest release
ROUTE 2

Ask your coding agent

Paste the official install request into Claude Code, Codex, or another coding agent.

PASTE THIS
Install AgentSmith for this project. Follow the official agent guide at
https://github.com/PromptPartner/agentsmith/blob/master/AGENT-INSTALL.md
Inspect the project first, explain your profile recommendation in plain language, show me the exact installation plan, and ask once before applying it.

Setup inspects the project, recommends a work profile from repository evidence, and shows the plan before applying it.

Advanced: set up from source

Building AgentSmith from source is for contributors and advanced automation. The source route requires Python 3.11+ and Git.

SOURCE SETUP
git clone https://github.com/PromptPartner/agentsmith.git
cd agentsmith
python3 agentsmith.py

The disposable First Verified Loop demo also runs from the source checkout.

DEMO FROM SOURCE
python3 agentsmith.py demo first-loop --target /tmp/agentsmith-first-loop

The demo prints the next steps and exact cleanup boundary.

Installation reference

05BOUNDED AUTONOMY

Delegate with boundaries.

Once your checks are trustworthy, approve a defined task and set explicit limits. A maker agent implements the change, configured verification runs, and a separate checker reviews the result. The run finishes with an accepted local result or an escalation. Publishing and merging require separate authority.

Coordinate independent approved tasks and verify their combined result before dependent work continues.

Finite runs guide Parallel work graphs
  1. 01Approved taskDefined scope and explicit limits
  2. 02Maker agentImplements the change
  3. 03Configured verificationYour project's checks run
  4. 04Separate checkerReviews the result

Accepted local resultorEscalation

06AGENT SUPPORT

Clear about what is verified.

The registry separates instruction discovery, skills and tools, and native runtime support. A target is not a passed certification.

NATIVE · PASSED

Claude Code + Codex

AgentSmith configures their documented local surfaces. Dated native-client evaluation records are available.

TARGETS · PENDING

14 other agent clients

GitHub Copilot, Cursor, Gemini CLI, Windsurf / Devin, Cline, Roo Code, Aider, Continue, OpenHands, Goose, OpenCode, JetBrains Junie, Zed, and Jules have adapters and registry-contract fixtures only. Client behavior is not certified; capability gaps are marked in the registry.

See the authoritative support registry

07START SMALL

Try one verified task.

The disposable demo refuses a non-empty target, guides the red-to-green task, and gives you the cleanup boundary. After that, set up AgentSmith in your own project.

Follow the guided first hour Get started for free