Version 0.1.0 for macOS and Windows
FireMoon Harness reads your code, edits it, runs it and proves the work is done. Bring any provider or a local model, keep your data on your machine, and pay once if you want more.
Free for macOS and Windows No account No telemetry
A real recording: a free model on OpenRouter fixes a failing test in a sandbox project. The waiting is sped up 3×; it cost $0.00.
Takeover mode lets the agent see and operate your desktop apps. It tests the app it just built, clicks through a UI, and finishes the task for you.
Illustration. Takeover footage will be here when it ships.
Why FireMoon
Everything below ships today, in a native app that starts fast and stays light.
Anthropic, OpenAI, Gemini, DeepSeek, OpenRouter, Mistral, Moonshot and Z.ai, plus local Ollama and LM Studio. No provider is the default; all of them are first class.
Agent mode is a full coding agent. Chat mode is a system assistant that can inspect and operate your computer. One click in the titlebar switches between them.
Background subagents work in their own git worktrees and branches. Nothing merges until you say so.
Four clear access modes, rules you control, and an undo for everything the agent did.
No telemetry. Free needs no account. Your sessions stay on your disk, and voice and image reading run on your machine.
The iOS companion drives your desktop sessions, answers approvals, and runs a chat right on the phone.
Run Claude Code on your Claude subscription, with FireMoon's interface on top. It needs Claude Code installed and signed in.
Providers
Bring your own keys, or your own hardware. Switch per session, per turn, or let Auto routing choose.
Compare
FireMoon next to the agents you may already use: models, privacy, price and the engine underneath.
| FireMoon Harness | Claude Code / Desktop | Codex | Cursor | Antigravity | DeepSeek Harness | |
|---|---|---|---|---|---|---|
| Models and providers | Every major cloud provider as a peer, with Auto routing and fallback | Claude models only | OpenAI by default; custom providers through config | Several vendors' models; own keys for chat only | Gemini plus a few partner models; no own keys | DeepSeek first; others through custom endpoints |
| Local models | Ollama and LM Studio, local or on another machine | Not supported | Ollama or LM Studio (--oss) | No official support | Through the SDK only | Any compatible endpoint; no presets |
| Context and cost savings | Caching, compaction, context slicing, cheaper-model routing, a cost meter with budgets, local image reading | Caching, auto-compaction, cost view, spend limits | Auto-compaction, token usage view | Auto-summarization, usage dashboard, spend cap | Auto-compaction, context panel | Auto-compaction, token and cache meter |
| Price | Free tier. Pro is €20 once. Your own keys or subscription | Subscription from $20/month, or API billing | Free tier; paid plans from $8/month, or API | Free tier; paid plans from $8/month | Free tier; Google AI plans | Free; you pay providers |
| Telemetry | None. Only update checks and licence validation | On by default; opt out with an env var | Anonymous analytics on by default; opt out in config | Privacy Mode available | Telemetry toggle in settings | Desktop analytics on by default |
| In-app notes | Git-native project notes for you and the agent | Instruction files only | Instruction files only | Rules files only | Rules files only | Instruction files only |
| Command palette | Every action, with aliases and shortcuts (Mod+K) | Slash commands | Command menu in the app | Editor command palette | Command palette in the app | Session search and slash commands |
| Interface | A native frosted-glass window with a custom titlebar | Terminal, desktop app, IDE extensions, web | Terminal, desktop app, IDE extension | A VS Code-based editor | Agent app, IDE and CLI | Desktop app and web UI |
| Engine | Rust core in Tauri. No Electron, no Node | Native CLI; Electron desktop app | Rust CLI; Electron app | Electron (VS Code fork) | VS Code fork (IDE) | TypeScript on Node.js, Electron |
| Platforms and companion | Windows and macOS (Linux soon), an iOS companion, sync through your own database | macOS, Windows, Linux; mobile remote control | macOS, Windows, Linux preview; phone via ChatGPT | macOS, Windows, Linux; iOS app for agents | macOS, Windows, Linux; browser remote control | Windows and macOS; no companion |
Competitor details checked against each product's own documentation and pricing on 11 November 2026. Products change quickly; if something here is out of date, tell us.
Use cases
Build, review and ship with any model, in parallel.
“Build the OAuth flow while a subagent fixes the flaky test.”
Schedules, research and your own infrastructure.
“Tidy my Downloads folder every night.”
Chat mode looks first, then asks before it changes anything.
“Why is my disk full?”
The boring jobs, done in seconds.
“Convert this folder of documents to PDF.”
Pricing
No subscription needed for the desktop. You bring your own keys or subscription.
€20one-time, lifetime updates
Everything in Free, plus the features that do more on their own.
Download
Version 0.1.0 for macOS and Windows. Linux is on the way.
FAQ
For cloud models, yes: you bring your own key from the provider, and you pay the provider directly. Keys stay on your computer, in FireMoon's credential store, unless you turn on encrypted key sync to your own database. Local models through Ollama or LM Studio need no key.
Yes, through Claude Mode (Pro). It runs Claude Code on your Claude account, with FireMoon's interface on top. Claude Code needs to be installed and signed in.
Yes. Point FireMoon at a local Ollama or LM Studio, or at Ollama on another machine of yours, and everything stays on your hardware.
Yes. The cost meter tracks spending per turn, conversation, day and model, computed locally, and you can set budgets. Subscription usage is kept apart from pay-as-you-go spending.
Only licence validation, and only for Pro and Max. Besides that, the app reads a public release file on GitHub to check for updates, which you can turn off. There is no telemetry, no usage data, and no session content sent anywhere but the model provider you chose.
On your disk, in FireMoon's data folder. Nothing is uploaded unless you turn on sync.
You bring your own Postgres database, and FireMoon syncs your devices through it. No FireMoon server is involved.
No. Speech is recognised on your computer, and no audio leaves it.
Not without asking. New sessions start in Workspace Write: the agent can change files in your project and asks before anything wider. Chat mode starts in Read Only.
Read Only can look around and asks before every change. Workspace Write can edit inside the project. Auto (Pro) has a background reviewer check every tool call and block dangerous ones. Full access runs without approval prompts.
Yes. Rewind restores the code, the conversation, or both to how they were before any message.
It starts read-only: it looks around, explains what it found, and asks before it changes anything.
