- Category
- AI Agents
- Rank
- No. 1190Tools index
Previous survey · No. 1147 ·
- Pricing
- Open Source
- Type
- AGENT
- Use case
- Coding
- Interfaces
- CLI · Desktop
- GitHub
- 255 stars
- Latest release
- supervisor-v0.1.1
- Date
About
Mixdog is a coding-agent CLI/TUI with a desktop app that lets users choose models, assign agent roles, and organize parallel sessions with tabs and split panes. It uses cache-aware context, focused tools and compaction to reduce token overhead, and the repository publishes same-model Terminal-Bench comparisons.
What it does
Mixdog is an open-source coding agent you run as a terminal program or a signed desktop app on Windows, macOS and Linux. You sign in with an existing Claude, ChatGPT, Grok or Muse Code subscription, or paste API keys, and it edits code, runs commands and searches your project. Its pitch is spending fewer tokens per task: it keeps a stable prompt prefix so provider caches hit, returns trimmed tool output, stores long history in a local database instead of the prompt, and uses a small local model to pick how hard the main model should think. It also drives browsers and desktops, and works on remote machines over SSH.
Why it's ranked here
The token claim comes with receipts. Across all 89 Terminal-Bench 2.1 tasks, five trials each on the same GPT model, it solved 385 runs to Codex CLI's 383 at 39% lower cost and 46% less final context; a single-pass run against Claude Code on Claude Opus 5 shows two more tasks solved at 19% lower cost. The project says every log and script is published. Add an Apache 2.0 license, a code-signed installer for every major desktop OS and a long list of supported providers, and it is a serious open alternative to the vendor agents. The caveat: the benchmarks are self-run, and one developer builds it.
What's good
The efficiency story is specific rather than vague. The context design doc lists ten separate mechanisms and is candid about limits: cached tokens still fill the window, and compaction reports an error rather than silently dropping required context. Code navigation runs on a native Rust engine with tree-sitter and ast-grep that parses 31 languages, so the agent asks for outlines and callers instead of rereading files. The SSH mode keeps your model logins on your own machine, listens only on the remote loopback behind a token, and turns off agent and X11 forwarding. Licensing is tidy, with a notice file that traces borrowed code back to its upstream license.
Tradeoffs
The scope is enormous: Office documents, image and video generation, email, schedules, phone control and chat bots ship alongside the coding agent, and the npm install pulls in a headless browser driver, a Postgres client, OCR, ONNX runtime and image processing, with install scripts that run on setup. The terminal UI library installs from a patched tarball on the project's own releases rather than the npm registry. The local model provider only works on Windows with an NVIDIA GPU. The SSH guide is written in Korean with only an English summary. The headline 63% figure is not one of the numbers in the published benchmark history table.
How to use it well
Pick it if token or subscription spend is your real constraint and you want one agent that can route work across several model vendors you already pay for. Start with the headless exec mode in scripts and CI-like jobs: it runs one model with no memory, skills or plugins, which makes runs reproducible, but note it is not a network sandbox. Keep the optional Cursor, Antigravity and MiMo sign-ins switched off, since the project itself warns they risk account restrictions. If you only want a lean terminal coding assistant, the sheer breadth here is more surface than you need.
Technical notes+
package.json publishes the npm package mixdog with bin src/cli.mjs, engines node ^22.19.0 || >=24.0.0, and os darwin/linux/win32 on x64/arm64; a postinstall script prunes the embedding runtime and prepares native assets, prepublishOnly blocks local publishes in favour of a GitHub workflow release, and ink resolves to a tarball on the project's GitHub releases. src/cli.mjs enables the V8 compile cache, performs a staged-update swap and re-execs a fresh process guarded by MIXDOG_SWAP_REEXEC, then hands off to run() in src/app.mjs, which routes help, the legacy readline REPL (--plain), headless exec through a pristine-route validator, or the bundled Ink TUI, warning when the TUI bundle is older than its sources. src/runtime/memory/index.mjs is a separate long-lived service that opens a lock file in the data dir, binds loopback ports 3350 to 3357, exposes an MCP stdio transport, caches embedding dimensions in a metadata file in the data dir to keep model warmup off boot, and delegates storage to a supervised Postgres. docs/context-efficiency.md maps each of its ten layers to implementation modules; NOTICE.md records the Codex CLI apply_patch Lark grammar carried verbatim, the effort judge as an 18-layer, 4-bit ONNX derivative of EmbeddingGemma 2 downloaded from releases, and a GPL-3.0 Chrome password import sidecar kept as a separate process. docs/computer-use-architecture.md describes C# and PowerShell host sources embedded into the desktop builds.
Observed
- License
- Apache-2.0 (LICENSE), third-party notices in NOTICE.md
- Language
- JavaScript ES modules on Node.js, with native Rust components
- Runtime requirement
- Node.js ^22.19.0 or >=24.0.0 (package.json engines)
- Install surface
- npm package mixdog; desktop installers for Windows x64, macOS arm64 and x64, Linux x64 and ARM64 AppImage and .deb
- Interfaces
- Interactive terminal UI, desktop app, headless exec command, MCP server for the memory service
- Model providers
- Anthropic, OpenAI, Google Gemini, xAI, Muse Code, OpenRouter, DeepSeek, OpenCode Go, and a built-in local provider
- Local model provider platform
- Windows x64 with an NVIDIA GPU only
- Long-term memory store
- Managed local PostgreSQL with pgvector and full-text indexes
- Remote servers
- SSH to glibc Linux or macOS hosts with Node.js 22+; Windows and musl Linux servers unsupported
- Benchmark records
- Self-run Terminal-Bench 2.1 comparisons against Codex CLI and Claude Code, scored by the Harbor verifier
Read from README.md, package.json, LICENSE, NOTICE.md, src/cli.mjs, src/app.mjs, src/runtime/memory/index.mjs, docs/context-efficiency.md, docs/computer-use-architecture.md, docs/ssh-remote.md.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.