winget install --id=Outcasts.Fuseforks -e Fuseforks is a desktop multi-agent AI orchestration tool. Connect AI agents into collaborative workflows, delegate tasks in parallel, gather results, automate recurring work, and integrate MCP tools - all from a unified interface. Supports major cloud and local LLM providers.
Outcasts Fuseforks
Keep a village of AI agents in your hands.
Outcasts Fuseforks is a desktop application for multi-agent orchestration, where multiple AI agents coordinate and converse with each other. Create agents, connect them, and talk to them — the village springs to life. They delegate, divide work, bundle results, and work on their own when the time comes. Everything is visible in a single screen across 3 panes.
Rust (fuseforks-core) + Tauri v2 + Vue 3 + Bun. The in-app display name is "Fuseforks."
| 🏘️ Build a Village | Create agents and tie them together. Kizuna is your control panel |
| 🤝 Delegation and Convergence | The coordinator asks with ask and distributes work to workers in parallel with plan, then bundles the results. Optional plan review (off by default) lets you edit the plan before it fans out (a schedule option or the status bar switch skips it for unattended runs). A default verifier appends a verification verdict to each bundle. The wait for an answer is configurable (default 600 s); circular delegation is rejected instantly instead of waiting |
| ⏰ Scheduling | Requests fire at times like "every Thursday at 17:00" or "every 10 minutes." No cron syntax required |
| 🔎 Pre-check | Run a command at firing time and ask only when its output matches your signal. Runs that do not match cost no tokens at all. Commands that arrive with a shared village never run until approved |
| ✅ Acceptance check | When the request finishes, run a verification command; if its output does not match your signal, the request is re-issued with the failure attached (up to an attempt limit — one retry by default). The check itself costs no tokens |
| 🔌 MCP | Paste Claude Desktop's mcp.json as-is. Shared + per-agent. Both stdio and remote (Streamable HTTP) |
| 🔍 Grounding | Gemini's Google Search and URL fetch, Grok's Live Search (web / X), web search on OpenAI and Meta, and Perplexity's web / finance / people search and URL fetch. The display distinguishes between searched facts, their sources, and facts that went unfound |
| 🧠 Thinking summary | The model's own summary of its reasoning, collapsed in a frame separate from the answer. Sources are verifiable pointers; a summary is an unverifiable claim — they are not mixed |
| 🛠️ Built-in Tools | remember / grep / fd / diff / sd / yq / file / rag / run. File tools are structurally unable to read outside the work folder (the exceptions are rag, which reads declared folders, and run, whose enclosure is its allowlist) |
| 🪧 Tool Reasons | When a servant reaches for a tool, what it is reaching for it for appears as one line in the conversation. It is the model's own account, not an audit record |
| 🎚️ Handoff toggle | Deny a coordinator the tool that passes the conversation on. Delegation (ask and receive) and splitting work remain, so answers come back instead of drifting to the user |
| 🗣️ Public Square Log | A village where you can hear others' conversations. You're also free not to listen (as a cost setting) |
| 📎 Path Completion | Type @ to pick a file from the work folder. Only the path is inserted, and the rounds a servant spends searching disappear |
| 🖼️ Attachments (image / audio / video / PDF) | Paste or pick a file in the input box and the addressed servant looks at it or listens to it. It reaches the model on that turn only (so the sliding window never resends it). Wires that cannot carry a kind warn on paste and refuse to send |
| 🏛️ Village Ordinance | Common rules that appear at the top of every agent's prompt. A normalization layer that unifies constitutional differences between models |
| 🎭 Roles | Templates for servants. Pick one at creation and the settings come with it; a colored badge shows in the list and in Kizuna |
| 🗂️ Groups | Section the servant list by task. Each heading toggles visibility and batch start; drag a card into another section to move it. Hiding changes only what you see — delegation and handoff ties are unchanged (Spec 51) |
| 📁 Change work folders together | When you point the whole village at another project, set every checked servant's work folder in one go. Running servants pick it up from their next message |
| 💾 Conversation Persistence | Close and reopen to pick up where you left off. Hold multiple conversations, switch between them, and fork from any point |
| 📊 Stats | What this village has paid, by conversation × servant × model × how each turn ended (a servant that switched models gets one row per model). Units are tokens (the same effective-token weights the budget uses). Register per-model rates and it also shows an approximate cost (≈ $) (Spec 41; rates come from a "fetch" button or by hand — the button also fills in the context length (Spec 50). Models without rates drop out of the total, and the screen says so). Failed turns are counted too. "All conversations" shows one month at a time, cut at a closing day, and you can page back to earlier months (the closing day is set under System Settings > Cost Management; the default is end of month, and "All time" is still available — Spec 42). Recording starts with this version (earlier conversations cannot be counted) |
| ⚙️ System Settings | Your own name and icon, language (switches both the screen and what the core says to the servants), token limit, confirmation dialogs. The left menu is the catalog of what can be configured |
The connection target is OpenAI-compatible / Anthropic / Gemini / xAI / OpenAI / Meta / Perplexity native. The base URL is flexible, so it connects directly to local LLMs like Ollama or LM Studio.
The intended users of this app are engineers as a hobby. We are not targeting orchestration infrastructure for business — in business there is accounting for labor costs, so "let AI do it instead of asking people to check" is justified, but for individuals the API cost catches the eye more than one's own time. That asymmetry cannot be solved from the app side.
Yet precisely because this is for a hobby, the insides must be real. A simple group chat tool is not worth an engineer's hobby time. Early Linux was dismissed as a toy by Solaris users, but because the internals were real Unix, it was worth taking home — that is the form we pursue.
Therefore, the design has 2 layers with different disciplines:
fuseforks-core / data_contract.yaml / firing rules) is production quality.
Freeze the contract before implementation, write tests red-first.
Zero dependency on GUI (guaranteed mechanically. This crate alone runs headless)Doing both half-way is the only failure mode. Don't loosen the contract for cuteness. Don't add configuration to look businesslike.
| OS | How |
|---|---|
| Windows | winget install --id Outcasts.Fuseforks — update with winget upgrade --id Outcasts.Fuseforks |
| macOS | brew install --cask betyourluck/tap/fuseforks (Apple Silicon only, signed & notarized; the fully-qualified name is required) |
| Linux / direct | .msi / -setup.exe / .dmg / .AppImage / .deb / .rpm from Releases |
Requirements: Rust 1.85+ (edition 2024), Bun, and the Tauri v2 prerequisites for your OS (WebView2 on Windows, WebKitGTK on Linux, Xcode CLT on macOS).
cd apps/gui-tauri && bun install
Run in development (with HMR):
cd apps/gui-tauri && bun run tauri dev
Build the distributable. Installers land in target/release/bundle/:
cd apps/gui-tauri && bun run tauri build
Tests and lint:
cargo test --workspace
cargo clippy --workspace --all-targets -- -D warnings
cd apps/gui-tauri && bun run test
> cargo test --workspace fails while the app is running — the executable cannot be
> replaced. cargo test -p fuseforks-core covers the core and works with the app running.
Pushing a v*.* tag runs release builds for 3 operating systems on GitHub Actions
(.github/workflows/build.yml). Ordinary commits do not
trigger it. The macOS build is Apple Silicon only — Intel Macs are not a target.
Core (crates/fuseforks-core)
Shell (apps/gui-tauri)
| DETAIL_en.md | Directory structure, concurrency model, screen layout, tool safety boundaries, LLM wire layer, operation |
| data_contract.yaml | The domain contract. It takes precedence over the implementation |
| specs/ | Specifications. Drafted, reviewed, then implemented in phases |
| failures.md | Traps stepped in (symptom → root cause → prescription → generalization) |
| PRIVACY_en.md | Privacy policy (the developer receives nothing) |
> The two documents above are written in Japanese.
MPL-2.0 (LICENSE). Why this license (2026-08-05):