AgentSmith is a shell-based harness that assembles a universal CLAUDE.md-backed operating core with swappable work profiles, enabling autonomous agent workflows across various tasks. It provides a setup.sh workflow, per-project profiles, and safety modes, with tooling to verify and deploy work items.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
AgentSmith is a universal, model-agnostic operating harness for AI agents (Claude, Codex, Gemini, …) built as a lean core plus swappable work-type profiles. It installs a CLAUDE.md-based harness around a fixed core and per-project profiles, allowing automated planning, implementation, verification, and shipping for tasks across software, devops, marketing, document creation, data crunching, and more. It includes scripts and documentation to assemble and configure the harness per project or globally.
How it works
The project presents a four-step model: 1) Install once to assemble a lean CLAUDE.md from the universal core plus profiles, 2) Run sessions that execute plan → implement → verify → ship with autonomous routing and minimal human intervention, 3) Use verify.sh to run project-specific checks, and 4) Hand off and improve the harness over time. The setup flow can be run via a wizard (setup.sh) or by passing flags to customize profiles, safety, and targets. The harness supports global core installation plus per-project profiles, and can self-update to incorporate core or profile changes. It includes a bundled skill pack and a variety of flags to control plugins, hooks, and behaviors.
Getting started
Getting started offerings include:
-
- Use the wizard by running the setup script:
./setup.sh(no flags = the wizard). It prints the command it will run and then executes after confirmation.
- Use the wizard by running the setup script:
-
- Run the wizard non-interactively by cloning and executing
./setup.shwith flags to specify profile, operator, target, and safety settings. Example commands are provided in the README, includinggit clone https://github.com/PromptPartner/agentsmith.git ~/tools/agentsmith && cd ~/tools/agentsmithfollowed by./setup.sh.
- Run the wizard non-interactively by cloning and executing
-
- Drive flags yourself by running
./setup.shwith options such as--profile software-dev,--operator-name,--operator-role,--tracker, and--target. Layered setup allows a global core install and per-project profile.
- Drive flags yourself by running
The repository documents Windows equivalents and a self-update flow via --self-update, as well as instructions to install from a zip or via git.
Recent releases
There are no releases listed in the repository metadata beyond the initial state, and the README does not indicate published release notes.
Traction
Stars: 307. Forks: 15. Open issues: 0. Language: Shell. License: MIT. Created: 2026-07-16. Last push: 2026-07-18.
Caveats
License: MIT. Requirements include Claude Code or another agent that reads CLAUDE.md/AGENTS.md. The setup relies on bash and jq (jq is optional but recommended). The README mentions Windows PowerShell port (setup.ps1) with PowerShell 7+ as a requirement. The latest releases section lists none, and there are no open issues.






