AgentSmith
Give coding agents project rules and verifiable completion checks, so changes come with evidence.
- Source repo
- PromptPartner/agentsmith
- Stars
- ★ 152
- Last updated
- 13d ago
- License
- MIT
- Primary language
- Python
- FA score
- 75/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Developers who want consistent project rules and completion checks for Claude Code or Codex across repositories.
- Not a fit if
- Teams expecting AgentSmith to choose or generate an application framework
- Teams requiring project code to run during setup inspection
- Source review
- 75/100 · Good
What does this agent do, and when should you use it?
AgentSmith is an open-source CLI for configuring project rules and work-type profiles for coding agents such as Claude Code and Codex. Its guided wizard inspects an existing project or initializes an empty folder, then previews the installation plan before writing files. The installed files include AGENTS.md as the canonical agreement, a generated CLAUDE.md that imports it for Claude Code, .harness/verify.conf for project checks, and .agentsmith/state.json for AgentSmith-owned settings. An agent uses the selected profile to make bounded changes, run configured checks, and retain verification evidence; decisions and handoffs can be saved in the project. Standalone builds are available for macOS, Windows, and Linux, include their runtime, and do not require Python.
Run agentsmith to start the setup wizard and choose an existing project or an empty folder to initialize with Git. The wizard inspects file evidence, recommends a work profile, previews the files it will write, and then adds project rules and verification scaffolding; inspection does not run project code or replace existing project configuration. The software-dev profile lists build, checks, tests, a security pass, and a real invocation as its main proof. Other profiles cover DevOps, marketing, documents, data, administration, research, design, and security audits. Optional advanced settings provide skills, MCP servers, and hooks. The agentsmith memory search and agentsmith memory read commands search or read local notes, while agentsmith audit-config produces a read-only advisory report.
- Developers who want consistent project rules and completion checks for Claude Code or Codex across repositories.
- Teams that want to inspect an existing project and preview a plan before adding verification scaffolding.
- Engineering teams that need build, test, security-review, and real-invocation evidence for features, fixes, or refactors.
- Maintainers who want agents to save decisions and a handoff so work can continue in a later session.
- Cross-functional teams that want work-specific completion criteria for operations, documents, research, design, and other tasks.
How do you install or deploy this agent?
Download the standalone build for macOS, Windows, or Linux from the latest release. The macOS and Windows builds are currently unsigned, so the operating system may ask you to confirm before running them. The build includes its runtime; Python is not required. Start the wizard with:
agentsmithAlternatively, ask Claude Code, Codex, or another coding agent to follow the official installation guide. The agent is instructed to inspect the repository, explain its recommendation, and show the plan before applying it. Contributors and advanced automation users can see INSTALL.md for source builds.
How do you use this agent?
Run agentsmith and answer the guided questions. Choose an existing project to configure, or provide an empty folder for a new project; the wizard shows the exact plan before writing. After setup, the agent follows the agreement in AGENTS.md and the checks in .harness/verify.conf. Optional local-memory commands include:
agentsmith memory search "topic"
agentsmith memory read docs/research/note.mdFor a read-only configuration audit, run:
agentsmith audit-configThe --include-user option adds native user settings, and --fail-on high adds a CI gate.
What are this agent's strengths and limitations?
- AGENTS.md is the single canonical agreement, and Claude Code gets a generated CLAUDE.md that imports it.
- The wizard previews its plan before writing and recommends a profile based on project file evidence.
- Project checks are defined in .harness/verify.conf; the example evidence bundle shows a failing test, verification after the fix, a real command invocation, a receipt, and a handoff.
- AgentSmith records only its own settings, helping updates and removal preserve foreign content.
- Optional profiles cover work beyond software development, including operations, documents, research, and design.
- The standalone macOS and Windows builds are unsigned, so running them may require user confirmation.
- The new-project flow initializes an empty folder and adds scaffolding; it does not choose or generate an application framework.
- Existing-project inspection does not execute project code, so configured checks must be run later to verify the project.
- The effects of optional skills, MCP servers, and hooks need to be reviewed separately before installation.
How does this agent compare with similar options?
The README positions AgentSmith as a way to move from vibe coding to evidence-based agent engineering, but does not provide a direct feature comparison with another product.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AgentSmith This agent | 75 · Good | CLIFree | ★ 152 | 13d ago | Python | Codex · Claude Code |
| Spec Kit | 72 · Some gaps | CLI | ★ 149k | 1d ago | Python | — |
| Keep the Why | 83 · Good | Agent plugin / skillFree | ★ 235 | 2d ago | JavaScript | Codex · Claude Code |
| Swarm Orchestrator | 73 · Some gaps | — | ★ 114 | 2d ago | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Cautious permissions, explicit consent for external writes, staged update plans and rollback receipts, secret scanning, and redaction guidance support ordinary safeguards. Autonomous runs add an accepted-spec gate, explicit authorization, and a local sandbox gate. The packet mainly contains policy and documentation rather than the core installer implementation, so auditable detail about data flows and dependency sourcing is limited.
The docs describe cross-platform checks, fail-closed behavior, and concrete failure cases; the work-graph review also leaves intermittent failures open. In the supplied AGENTS.md, an agent is told to research and continue after ambiguity, while the autonomous-run skill requires explicit authorization and narrowly states that its controller has no external-write adapter. Their precedence and scope are not fully reconciled here, lowering self-consistency. No run or test was executed.
Profiles cover development, operations, research, administration, creative work, and security, and distinguish ordinary work from finite autonomous runs. The docs discuss container constraints and client-specific configuration. Skill triggers are generally specific, but only a sample of skills is supplied and some boundaries depend on client behavior and host sandbox support.
The repository materials show layered docs, installation guidance, troubleshooting, update notes, known limitations, a security policy, and an MIT license. Setup previews a plan and asks for confirmation; updates produce rollback receipts. The main installer source is not in this packet, the full release history is not shown, and individual maintenance responsibility is unclear.
The core asks agents to turn completion criteria into checks and evidence; the handoff skill specifies recovery material for a fresh session, and the example skill and demo tests illustrate user-visible outcomes. This offers clear value for evidence-driven work, while configuration, checking, and record-keeping add ongoing effort. The supplied efficacy claims are mainly documentation claims, and no evaluation was run.
Several claims point to named docs, checks, test files, and open issues. The configuration-audit doc distinguishes static pattern checks from execution evidence. The work-graph review describes commit/tree matching for reports and admits that declarations are not independent telemetry. Evidence is mostly from the same repository; the supplied material does not fully establish how some historical reports map to this revision, and many results remain maintainer-reported.
- This assessment uses only the supplied static sample; the full repository, executable, installer, and tests were not inspected or run.
- Autonomous runs require an accepted spec, implementation ticket, operator authorization, and available sandbox. The docs say cooperative resource locks do not replace operating-system isolation.
- The packet records unresolved intermittent macOS cleanup-preview and Windows child-controller failures. Unsigned macOS and Windows builds may prompt for confirmation.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/agentsmith)<a href="https://followagents.com/en/agents/agentsmith"><img src="https://followagents.com/badge/agentsmith.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.5.0 AgentSmith 0.5.0
Release notes excerpt
AgentSmith 0.5.0 makes the always-loaded rules about 40% shorter and closes the installer-safety issue #50. It also adds checks that show whether your rules actually take effect, and guards for unattended runs. Highlights - Leaner rules. The core and the software-dev profile drop from about 5,800 to about 3,350 words. Generic coding advice that current models follow unprompted is gone. Consent, secrets, identity, authority and the trust gates stay. The static-context budget is now 480 lines / ~7,800 tokens (#64). …View on GitHub ↗ -
v0.4.0 AgentSmith 0.4.0
Release notes excerpt
What's Changed * feat: make AgentSmith setup beginner-ready by @PromptPartner in https://github.com/PromptPartner/agentsmith/pull/39 * Explain AgentSmith through guardrails and proof in all five README languages by @PromptPartner in https://github.com/PromptPartner/agentsmith/pull/40 * Harden autonomous recovery and add offline qualification evidence by @PromptPartner in https://github.com/PromptPartner/agentsmith/pull/41 * Add optional local memory, hook disclosures and configuration audit by @PromptPartner in https://github.com/PromptPartner/agentsmith/pull/42 …View on GitHub ↗ -
v0.3.1 AgentSmith 0.3.1
Release notes excerpt
AgentSmith 0.3.1 adds safe, evidence-backed integration validation and hardens durable recovery. Highlights - A conditional MCP/integration checkpoint records discovery versus exact pins; configured, authenticated, read-tested, and least-privilege states; auto-start, capability listing, and actual invocation; named consumers and credential classes; response minimization; persistent state; shared-store causality; and exact child-process cleanup. …View on GitHub ↗
FAQ
Does AgentSmith cost money?
Will AgentSmith run my project during installation?
Will it overwrite my existing project configuration?
Can AgentSmith build a new application for me?
Can I run agents unattended?
Thanks — we use this to improve the page.