Description
AgentSmith is an open-source tool designed to transform the process of coding agents from a casual approach to a more structured engineering practice. It achieves this by implementing project rules, configured checks, and verification evidence, which help ensure that coding tasks are completed with precision and reliability. The tool is particularly compatible with Claude Code and Codex, allowing users to define project rules and a clear definition of done, which are then used to guide the coding process.
Read More
The tool operates on a local-first basis, meaning that it prioritizes local execution and storage, which enhances security and control over the coding process. AgentSmith is MIT licensed, making it freely available for modification and distribution, although users should be aware that any subscription or model usage costs associated with Claude Code or Codex are separate.
AgentSmith's workflow involves understanding and bounding tasks by reading existing code and defining scope and completion checks. It emphasizes making small, reviewable changes and explicitly choosing native permissions. The tool also includes optional secret-scanning hooks to enhance security.
Verification is a key component of AgentSmith's process. Users are encouraged to reproduce failures before fixing them, run all configured project checks, and exercise user-visible paths. The results of these checks are recorded in a receipt, providing a clear record of what has been verified.
AgentSmith also supports continued improvement by saving decisions and providing a clear handoff for future sessions. Observed failures can be captured and used to propose new rules, hooks, or tests, which are validated before being adopted. This iterative process helps ensure that the coding harness is continually improving and adapting to new challenges.
- CategoryAI Agent Builders
- Pricing ModelOpen Source
- DeploymentLocal
- PlatformsmacOS, Windows, Linux
AgentSmith's Core Features
Open-source harness for coding agents
Supports Claude Code and Codex
Local-first approach
MIT licensed
Project rules and configured checks
Verification evidence and handoffs
Optional secret-scanning hooks
Iterative improvement process
How to use AgentSmith?
Download: Obtain the standalone build for your OS
Run: Open a terminal and execute agentsmith
Configure: Answer guided questions to set up
Use: Implement project rules and checks
AgentSmith's Use Cases
- Project Rule Implementation
- Verification Process
- Iterative Improvement
- Local Execution
- Agent Compatibility
