fosnie

Fosnie is an open-source, self-hosted private AI platform. This repository is the open-source (Apache-2.0) edition; Fosnie Enterprise is the commercial tier (one product, one tier). It runs entirely inside your own infrastructure: no data, prompts, or documents leave it unless you explicitly enable an external connector. Bring your own model, a local LLM (vLLM, Ollama, llama.cpp) or an external API key (Anthropic, OpenAI, Gemini), and Fosnie gives you retrieval-augmented chat, agents, document work, research, and voice on top, in the browser, on your desktop, and on your telephone.

Built for teams that hold other people's confidential data under a duty of care (legal, audit, finance, healthcare), where "where does my data go?" is the first question.

Features

• 🔒 Offline by default: no data, prompt, or document leaves your infrastructure unless you enable a connector; every outbound connector ships dormant and opt-in.
• 🧠 Agentic RAG: decompose → hybrid (dense + sparse) search → rerank → grade → answer, with inline citations.
• 🔌 Model-agnostic + BYOK: point each capability (LLM, embeddings, rerank, OCR, STT/TTS) at a local engine or any external API, configured in the UI. Native Anthropic adapter (tools + extended thinking).
• 🧭 Configurable model routing: build Custom Auto and Full Auto profiles with editable presets, task and complexity rules, and model or reasoning-effort assignments. Personal settings can use shared models without a personal API key. Capability checks, locality controls, shared usage limits and bounded technical fallback are built in. Included in Core and Enterprise; see the routing profiles guide.
• 🤖 Agents & workflows: action-taking agent runs with human-in-the-loop approval, durable resume, and event-driven automations.
• 🔎 Deep Research: multi-step research over your documents, the web, or both, with a fully cited report.
• 📄 Document work: generate DOCX/PPTX/PDF/XLSX/HTML, tracked-change accept/reject, tabular review.
• ☎️ Telephone: point a number at your instance and an agent answers it, taking messages, screening callers, booking appointments and handing over to a person. Carried by a telephone company or by your own telephone system, where the audio never leaves your network.
• 🖥️ Desktop app: Windows and macOS, paired to your own instance, able to work in a folder on your computer with every write, deletion and command shown to you first and reversible.
• 🗣️ Voice: speech-to-text and text-to-speech, including live streaming that searches your Libraries while you speak.
• 🐍 Code interpreter: agents write and run real Python in a zero-egress sandbox built fresh for each run (Firecracker or gVisor).
• ✅ Groundedness: verify answers against their sources and flag unsupported claims.
• 🛠️ Tools & MCP: built-in tools plus one-click Model Context Protocol connections over OAuth 2.1.
• 🔗 OpenAI-compatible API: point your own code at your instance with the API client you already use.
• 👥 Teams & messaging: group and project chats and direct messages, with at-rest encryption.
• 🧾 Tamper-detecting audit: a SHA-256 hash-chained, append-only audit log.

Quickstart

Requirements: Docker + Docker Compose v2 (≥ 2.24). ~8 GB RAM with an external model API; ~16 GB for the fully-local stack (--local). Only one port is published (8080).

Option A: one-line install (downloads a pinned release, generates secrets, starts the stack):

curl -fsSL https://get.fosnie.dev | sh
# fully local: no API keys, all inference on this host:
curl -fsSL https://get.fosnie.dev | sh -s -- --local

Prefer to read before you run: download install.sh, review it, then sh install.sh.

Option B: manual (Docker Compose):

mkdir fosnie && cd fosnie
curl -fsSL -O https://github.com/Fosnie/fosnie/releases/latest/download/docker-compose.yml
curl -fsSL https://github.com/Fosnie/fosnie/releases/latest/download/example.env -o .env
# edit .env: set POSTGRES_PASSWORD, MESSAGE_ENCRYPTION_KEY (openssl rand -base64 32), ML_SHARED_SECRET
docker compose up -d

Open http://localhost:8080 and create the first account, which becomes the admin. In the default (external-inference) mode an onboarding checklist walks you through adding a model provider under Settings → Providers (a local engine or an API key). With --local the Ollama + reranker stack is wired up automatically, so chat works immediately.

• Fully-local models: --local adds Ollama (LLM + embeddings) and a llama.cpp reranker; the backend auto-configures them. First run pulls the models (a few GB).
• Managed Postgres (Supabase, Neon, RDS, Cloud SQL, Azure): set FOSNIE__DATABASE_URL in .env and drop the bundled postgres service (see the comments in example.env).
• Bare-metal (systemd), the offline air-gap bundle, backups, upgrades, and reverse-proxy TLS: see backend/deploy/README.md and the documentation.

Upgrade: docker compose pull && docker compose up -d (pin APP_VERSION in .env for reproducible upgrades).

Editions & Licensing

• Fosnie (this repository) is free and open-source under Apache-2.0, and is a complete, self-contained product: everything above, with nothing locked behind a paywall.
• Fosnie Enterprise is a separate commercial edition for larger and regulated organisations. It depends on Core and adds compliance, governance, and scale features. See the comparison below or talk to us.

Fosnie vs Fosnie Enterprise

Architecture

• backend/ is Rust (axum + tokio): chat orchestration, WebSocket transport, auth, scheduler, audit, config. sqlx compile-time-checked queries; forward-only migrations.
• ml/ is Python (FastAPI), the only LLM client: agentic RAG, extraction/OCR, chunking, embeddings, reranking, generation, voice, Deep Research, web search.
• frontend/ is a React 19 + Vite single-page app.

External inference, Qdrant, Redis, and PostgreSQL are reached as configured services; the platform is their client. Full map in the documentation.

Roadmap

See the roadmap.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md. We use a Developer Certificate of Origin sign-off on every commit (git commit -s) and follow our Code of Conduct. Security issues: SECURITY.md.

Community

Questions and ideas → GitHub Discussions · Bugs → Issues.

License

Fosnie is licensed under the Apache License 2.0. "Fosnie" and "Private AI" are trademarks of Private AI Ltd (see NOTICE); the licence does not grant rights to the name or branding.