Hey HN, I built Agent Smith — a self-hosted AI coding agent that takes a ticket reference, clones your repo, analyzes the code, writes an implementation plan, executes it, and opens a PR.
It supports GitHub, Azure DevOps, Jira, and GitLab. You bring your own API key — Claude, OpenAI, or Gemini. No SaaS, no account, runs on your machine or your cluster.
I built it in a few days using the same approach the agent itself uses: structured architecture prompts, strict coding principles, and an AI assistant doing the implementation. The coding principles that govern Agent Smith's output are the same ones I used to build it.
It's early — works well for well-scoped tickets, not yet reliable for large multi-file refactorings. Interactive chat interfaces (Slack, Teams) are in progress.
Would love feedback. The prompts and all 17 architecture phases are in the repo if you want to see how the context is structured.
Comments
hleichsenring•1h ago
Hey HN,
I built Agent Smith — a self-hosted AI coding agent that takes a ticket reference, clones your repo, analyzes the code, writes an implementation plan, executes it, and opens a PR.
It supports GitHub, Azure DevOps, Jira, and GitLab. You bring your own API key — Claude, OpenAI, or Gemini. No SaaS, no account, runs on your machine or your cluster.
I built it in a few days using the same approach the agent itself uses: structured architecture prompts, strict coding principles, and an AI assistant doing the implementation. The coding principles that govern Agent Smith's output are the same ones I used to build it.
It's early — works well for well-scoped tickets, not yet reliable for large multi-file refactorings. Interactive chat interfaces (Slack, Teams) are in progress.
Would love feedback. The prompts and all 17 architecture phases are in the repo if you want to see how the context is structured.
hleichsenring•1h ago
It supports GitHub, Azure DevOps, Jira, and GitLab. You bring your own API key — Claude, OpenAI, or Gemini. No SaaS, no account, runs on your machine or your cluster.
I built it in a few days using the same approach the agent itself uses: structured architecture prompts, strict coding principles, and an AI assistant doing the implementation. The coding principles that govern Agent Smith's output are the same ones I used to build it.
It's early — works well for well-scoped tickets, not yet reliable for large multi-file refactorings. Interactive chat interfaces (Slack, Teams) are in progress.
Would love feedback. The prompts and all 17 architecture phases are in the repo if you want to see how the context is structured.