Codewhale

Codewhale is the terminal agent for people who want to choose exactly how much rope it gets. Four approval modes, Plan, which is read-only, Ask, Auto-Review and Full Access, mean you can start with an agent that cannot touch a file and loosen it as trust builds. The trait that decides it is that graded trust; most agents give you allow-all or ask-everything.

Pick it if you run local models and like a small Rust binary. Pick OpenCode if you want a wider provider list and more people to ask when something breaks, and Aider if you want git as the safety rail.

The risk is identity. The project was deepseek-tui and became Codewhale with configuration and session compatibility preserved, so the codebase carries a former life aimed at one vendor, and the docs, packages and issues from that life still turn up under the old name. The maintainer is a single GitHub account, so an issue waits for one person.

Check which name your package manager installed. What it does right: unknown model prices stay unknown instead of being reported as free, a small honesty about the meter that larger tools do not manage.

Two documented properties. 1. Fleet mode coordinates several agents while keeping their internal instructions out of the user's transcript, which is context isolation stated as a design goal: the coordinator sees results, not the workers' prompts. 2. codewhale exec runs a task without the TUI, so the loop is invokable from a script and, in principle, reproducibly. Edits and commands are gated by the selected approval level.

No benchmark is published and none is claimed. The observation: hiding worker prompts from the transcript is good hygiene and also removes the evidence you would want when a worker goes wrong.

Hmbown is a maintainer, not a firm. The repository has 40,912 stars, an MIT licence and no pricing, no company page and no hosted tier, and the agent spends other vendors' tokens, so no revenue passes through it at all. That is fine for a project and unanswerable for an investor.

Moat: none; provider-neutral by design means no lock-in, in either direction. Likely outcome: the maintainer is hired, or the project fades when attention moves, or it stays a well-liked tool at this size. Likely acquirer: none. Position: no position; use it as long as it is maintained.

The demo is a Rust binary answering in a TUI. Procurement: there is no vendor, no seat and no contract, so the cost is sixty provider keys and whatever they spend, with no central cap. There is no SSO, no audit log and no support organisation beyond one maintainer's issue tracker. Shell completions ship for bash, zsh, fish, powershell and elvish, which is a small sign of care.

It runs non-interactively, so a CI job could call it. Onboarding is an npm install and a key. Approved with conditions: company-issued keys, sandboxed CI runners, and the least permissive mode on shared machines.

MIT and Rust, so I build it with Cargo, and the repo also ships Nix, Docker and Scoop routes. Local models are not an afterthought: Ollama, vLLM and SGLang are named backends, and the README's phrasing is connect a provider or stay offline, which is my preferred state. MCP servers attach, and the starting configuration is a config.example.toml I copy and edit.

The rough edge is documentation depth; the config file's location is not spelled out on the front page, so I read the source, which I was going to do anyway. Forkable in a weekend. Mine, with a whale on it.