KLINCH

An AI coding agent that works in your repository

Describe a task in plain language. The agent explores the code, makes exact edits, checks them and stops at a commit proposal that you review line by line.

What an AI coding agent does

A chat assistant answers questions. An agent acts: it calls tools in a loop until the task is done. In KLINCH the agent lists files, reads the ones it needs, searches the repository, edits code with exact replacements, creates new files and validates syntax. Every tool call is shown in the conversation, so you can follow its reasoning.

  • read_file and search_files: reads only what it needs
  • edit_file and write_file: exact, reviewable changes
  • run_command validate: syntax checks for JSON, TypeScript, JavaScript, CSS and HTML
  • propose_commit: a proposal, never a push

You stay in control of every commit

The agent cannot push. Changes are staged as a diff: you accept or reject each file, then KLINCH writes one commit for the accepted change set through the GitHub API. If your repository deploys on push, accepting a change ships it.

Bring the model you trust

KLINCH speaks the OpenAI Chat Completions protocol, so the agent runs on OpenAI, OpenRouter, DeepSeek, Kimi or any compatible gateway. You pay the provider directly; your key is encrypted and used only on the server.

Good first tasks

Agents shine on well-scoped changes. Start with something you can verify quickly in the diff.

  • Add a README with setup instructions
  • Find and fix TypeScript errors in a folder
  • Create a landing page or a new component
  • Explain the architecture of an unfamiliar repository

FAQ

Can the agent run tests or install packages?

No. KLINCH runs on Cloudflare Workers and in your browser, where there is no operating-system shell. The agent can validate syntax and read CI status; builds and tests run in your CI after the commit.

Does the agent see my whole repository?

It sees the file tree and reads files on demand through the GitHub API, only the ones the task needs. File contents are sent to the model provider you configured.