// the take
Codewhale is a Rust terminal coding agent (started life as deepseek-tui) that reads a repo, edits files, and runs shell commands to complete a task, working with 40+ model providers or local runtimes like Ollama/vLLM. It's aimed at developers who want an agentic CLI coding tool without being locked into one model vendor, and who care about auditability of what the agent actually did.
Provider-agnostic by design, not just an OpenAI wrapper: real support for local models and OpenAI-compatible endpoints alongside 40+ hosted providers, so you're not stuck if a vendor changes pricing or terms. The permission model (Ask/Auto-Review/Full Access) pairs with actual OS sandboxing (Seatbelt on macOS, bubblewrap on Linux) and specific protections against credential-store reads and Unicode-obfuscated commands — this is more thought-out than the typical 'y/n to run a command' approval flow. The receipts/undo system logs every file edit, command, and MCP/web call per turn and lets you roll back a whole request in one step, which is the right answer to 'what did the agent actually touch' that most agentic tools punt on. One engine backs five different interfaces (TUI, `exec`, web UI, Runtime API, ACP for editors like Zed), so it isn't a loose bundle of scripts.
Windows has no OS-level sandbox at all, and Full Access mode disables sandboxing everywhere else too — for a tool whose core feature is running arbitrary shell commands, that's a meaningfully weaker safety story than the approval UI implies. Telemetry is on by default (a rotating install ID plus usage/feature counts), which is fine for most people but something security-conscious teams will need to explicitly turn off and audit. The README includes a paid affiliate placement (a referral deal with a model router) baked right into an open-source project's front page — worth knowing monetization is already present before assuming this is purely community-driven. The codebase has sprawled into dozens of crates covering provider routing, cloud-fact overlays, catalog corrections, and more; for a tool that executes commands with your credentials, that's a large and growing attack/maintenance surface to trust long-term.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →