Coderrr

Every task begins with a written plan, not an edit. Coderrr reads your code, pulls relevant skills, and generates requirements.md, design.md, and tasks.md before doing anything else. The plan is shown to you, and the tool stops. This matters because you can correct course in plain English before any file changes, which is far cheaper than undoing a bad patch. It's the core loop that separates this terminal coding tool from agents that edit as they go.

The agent isn't asked to refrain from editing; it simply has no editing tool available while it plans. Write tools are absent from the model's tool list entirely, and approval is the event that unlocks them. It's a structural safeguard rather than a polite request, which is a meaningful difference when you're pointing an AI at a real repo. Structure beats promises.

Coderrr has no tool that runs commands against your working tree. Commands execute in a sandbox, and the agent reads the real exit code and output. No guessing. The default scratch tier uses a throwaway copy of the project, which limits the blast radius of a bad instruction. When Docker is available, it automatically upgrades to a container with networking off and all capabilities dropped. That upgrade matters, because a network-isolated container stops the agent from pulling in a dependency or pinging an endpoint you never approved during the run.

Plans live in .coderrr/specs/NNN-slug/ and are meant to be committed alongside your code. Beyond documentation, they double as the agent's memory: a later session reads tasks.md to learn where things stand instead of replaying a chat log. You can edit those files before approving to steer the work. That last part is underrated, since it lets you, for example, drop a task you don't want run today and keep a record of the decision inside the repo itself.

Coderrr reaches every provider over plain HTTP, so no provider SDKs get installed. It supports Ollama, Anthropic, OpenAI, Google, and OpenRouter. The default is gemma4:31b-cloud through Ollama, usable on a free account. API keys resolve from environment variables first, then the OS keyring, then a config file written with restricted permissions.

Skills are markdown guidance about how to approach a class of problem, fetched when the agent decides it needs them and deleted after use. They add no executable capability, so a poor skill can only give bad advice to an agent whose tools are already gated. You can browse the same registry yourself with coderrr skills search. That's a smart boundary, because it means an untrusted skill pulled from a public index can sway the agent's judgment without ever gaining the power to touch your disk.

You can connect your own MCP servers, and their tools become tools Coderrr can call: Figma, Linear, Notion, or something internal to your company. No extra install is needed, since the client is plain HTTP like every other connection in the tool. From inside a session, you paste a URL and you're set. Handy. For teams that already run a Model Context Protocol server for ticketing or design handoffs, this is the piece that folds Coderrr into work you're doing anyway rather than a separate island.