an OSS desktop app for multi-model AI dev.
Plan‑building Task conversations · Multi‑model code review · Self‑healing autonomous Swarms
Warning
Hyvemind is in alpha. It is being shared with friends and early testers.
Hyvemind is a desktop app that combines three modes of AI‑assisted development in a single GUI:
A focused conversational interface for building a plan. Every Task is a back‑and‑forth with an AI model of your choice, that ends in a workable plan you can hand off to an agent that will implement it, OR to a Hivemind which will strengthen the plan before implementation.
A concurrent multi‑model review engine. You define a team of LLMs and rounds. Each round runs N models in parallel against the same prompt that an Orchestrator puts together — based on the original plan, gathered source context, and rules. Outputs from a round are merged and fed into the next round, producing iterative refinement.
The Orchestrator will also score the hivemind reviewers and display the findings for you to get a personal feel of how well models do.
Fully autonomous multi‑feature execution. Hand the swarm a goal and a working directory; it runs until the work is done — Queen decomposes, Scouts plan, Workers implement, Guards validate, Nurse keeps it alive when things stall. Best of all, Hiveminds can be invoked at the Queen and Scout level. Swarm plans can be exported, cloned and used against different model compositions!
The combination is the moat. No single feature here is unique on its own — but no other product brings them all together at this quality bar.
| Differentiator | Why it matters |
|---|---|
| Self‑healing autonomy via Nurse | Other "leave‑it‑running" agents die on the first stall. Nurse runs seven detectors over every long‑running session and routes each signal through a three‑tier pipeline (deterministic → templated playbook → LLM classifier), so most interventions never even spend tokens. |
| Polished multi‑model review | Adversarial stance forces reviewers to find flaws. Cross‑round merge produces iteration, not just N opinions. |
| Full pipeline integrated | Tasks, multi‑model review, and autonomous swarms in one product, with shared cost tracking and a unified UI. |
| Built like real infrastructure | Circuit breakers, semaphore‑bounded concurrency, atomic file writes, OS‑keychain secrets, append‑only progress logs, crash‑safe state. |
| Premium boutique craft | The bar is Linear / Arc / Raycast — opinionated, polished, feels expensive in a good way. |
Every agent in Hyvemind maps to a real role in a bee colony. This is intentional. Each agent has its own system prompt and a specific thinking level.
| Agent | Role | Thinking | What it does |
|---|---|---|---|
| Queen | Orchestrator | Medium | Decomposes the goal into features + dependencies + milestones. Runs the swarm loop. |
| Scout | Planner | High | Analyses the codebase per‑feature, produces an implementation plan, complexity, risks. |
| Worker | Implementer | Medium | Writes the actual code following the Scout's plan. Emits a structured handoff JSON on completion. |
| Guard | Validator | Medium | Validates milestone assertions. On failure, synthesises a fix‑feature (max 3 retries). |
| Nurse | Heartbeat | Low (Tier 3 only) | Push‑mode supervisor for every long‑running session. Seven detectors (stall, reasoning loop, tool failure, process health, provider health, context saturation, retry exhaustion) feed a three‑tier dispatcher (deterministic action → templated playbook → LLM classifier). Decisions: LeaveIt / Steer / Restart / Cancel with mandatory kill verification. Every decision is logged to ~/.hyvemind/debug/nurse/ regardless of debug mode. |
- macOS primary (Intel + Apple Silicon). Linux & Windows builds are produced by CI but are not actively tested.
- Rust (stable) — install via rustup
- Node.js 20+ & npm
- Bun — for compiling the bundled Pi runtime
(
curl -fsSL https://bun.sh/install | bash) - An API key for at least one provider: Anthropic, OpenAI, OpenRouter, or one of the other providers we offer (see providers below).
git clone https://github.com/Unravl/Hyvemind.git
cd Hyvemind/app
npm install
npm run tauri:devThe first launch triggers a one-time Pi build (see First run below). On first run, open Settings to add your provider API keys — they're stored in your OS keychain, never on disk in plaintext.
Download the latest installer for your platform from Releases. Bundles are not yet code‑signed, so on macOS you'll need to right‑click the app → Open the first time to bypass Gatekeeper. Notarisation is on the roadmap.
The first npm run tauri:dev (or tauri:build) triggers scripts/build-pi.sh, which:
- Downloads the pinned version of Pi
(
@earendil-works/pi-coding-agent) viabun - Downloads Pi's bundled npm extensions (
pi-web-access,pi-subagents,pi-mcp-adapter) plus the two local Hyvemind extensions (hyvemind-providers,hyvemind-handofffromapp/src-tauri/pi-extensions/) - Compiles Pi into a single executable via
bun build --compile
This step can take 2–5 minutes on a first run — bun is downloading and compiling a
substantial TypeScript project with all its dependencies. The terminal will show
prepare-pi with no progress bar during this time.
The build writes a stamp file at app/src-tauri/binaries/.pi-version. As long as the
pinned Pi version in scripts/pi-version.txt hasn't changed, every subsequent
tauri:dev skips the build entirely — the stamp check costs milliseconds. To force a
rebuild (e.g. when upgrading Pi), bump the version string in scripts/pi-version.txt.
- Shell — Tauri 2 (Mac primary; Linux/Windows experimental)
- Frontend — React 18, Vite, TypeScript, Tailwind CSS, Vitest
- Backend — Rust (tokio, sqlx, reqwest, moka, keyring, tracing, sentry)
- Agent runtime — Pi by Mario Zechner /
earendil‑works, bundled and pinned (see
scripts/pi-version.txt) - LLM providers — Anthropic API, OpenAI API, Claude Subscription, ChatGPT Subscription, OpenRouter, OpenCode Go, CROF, Ollama, NeuralWatt, DeepSeek API, Xiaomi Mimo API, z.ai (GLM), NVIDIA NIM — and any OpenAI Completions compatible API. All behind per‑provider circuit breakers and a shared response cache
Architecture details live in CLAUDE.md. Product reference lives in
PRODUCT.md.
Topical deep-dives live under docs/. Start with docs/README.md
for the full index. Highlights:
docs/architecture.md— system component map + sequence diagramsdocs/frontend-architecture.md— React shell, provider tree, IPC layerdocs/bee-agents.md— per-agent deep-dive (Queen / Scout / Worker / Guard / Nurse)docs/ipc-reference.md— per-command Tauri IPC referencedocs/providers.md— LLM provider abstractiondocs/extension-authoring.md— Rust pollers, Pi extensions, topbar widgetsdocs/developer-cookbook.md— task-oriented recipes
PRs welcome — please read CONTRIBUTING.md first. By participating
you agree to abide by the Code of Conduct.
For deep‑dive architecture (file paths, IPC surface, debug commands, investigation
guides), see CLAUDE.md.
Found a security issue? Please don't open a public issue. See
SECURITY.md for the disclosure process.
Hyvemind is built on top of Pi
(@earendil-works/pi-coding-agent) by Mario Zechner / earendil‑works. Pi provides the
underlying agentic runtime; Hyvemind layers Queen / Scout / Worker / Guard / Nurse
orchestration, multi‑model review, cost tracking, and persistence on top. Pi is bundled
inside Hyvemind at a pinned version — users never install it themselves.
MIT © 2026 Hayden Evan