A self-hosted AI workspace for chat, agents, research, documents, email, notes, calendar, and local model workflows.

Quick Start · Setup Guide · Contributing · Roadmap

---

🤝 Odysseus × SAW — the Round Table

This is a community fork that combines Odysseus with the SAFe Agentic Workflow (SAW), built with Claude, to give Odysseus an easy-to-understand, end-to-end AI agentic workflow — the Round Table.

Instead of one chatbot guessing at an entire task, a small team of role-specialized AI agents passes the work down a real assembly line — Analyst → Architect → Developer → QA → Security → Tech Writer → Release — each reviewing the previous one's output behind quality gates. You write one ticket; the team plans, builds, tests, and documents working code right in your workspace.

What it brings to Odysseus

• 🧩 One ticket → working software — describe what you want and seven agents take it from spec to shipped code, no step-by-step prompting.
• 🚦 Quality gates, not vibes — a deterministic build/structure check (a Flutter app must really have a runnable lib/main.dart; Python must compile; flutter analyze must be clean) plus independent QA and Security reviews catch broken work before it can pass.
• 🎚️ Per-role model routing — run the Developer on a strong API model and the rest on free local Ollama models, or any mix, from the ⚙ Models panel — cost and privacy on your terms.
• 🔁 Continue the discussion — after a run, just ask for changes; the team remembers what it built and iterates on top instead of starting over.
• 🕘 History & reuse — every run is saved; replay any ticket or load it back into the form.
• 🔐 Least-privilege agents — non-coding roles can't execute commands, so a reviewer can't edit the code it's grading and a writer can't launch your app.
• 🤝 Human-in-the-loop releases — work commits to your branch locally, or opens a real GitHub PR — your call.

Open it from the ⊹ Round Table button in the left rail. Full walkthrough in the Round Table guide.

---

Quick Start

🪟 Windows — one-click installer (easiest)

Grab OdysseusSetup.exe from the latest release and run it. It installs a fully self-contained app (no Python, no Docker, no setup) with a Start Menu shortcut. For local AI models, install Ollama — the installer offers this if it's not found; cloud API models work without it.

🍎 macOS — drag-and-drop app (Apple Silicon)

Grab Odysseus-macOS-arm64.dmg from the latest release, open it, and drag Odysseus into Applications. The app is self-contained (no Python, no Docker).

It isn't notarized with Apple, so recent macOS versions block the first launch with a misleading "Odysseus is damaged" dialog (the app is fine — that's Gatekeeper-speak for "not notarized"). Unblock it once with:

xattr -cr /Applications/Odysseus.app

then open it normally. (Alternative: System Settings → Privacy & Security → Open Anyway.) For local AI models, install Ollama for macOS; cloud API models work without it.

🐳 Docker (any platform)

dev and saw-harness both track the newest changes in this fork.

git clone https://github.com/Z-Gamez/odysseus-roundtable.git
cd odysseus-roundtable
cp .env.example .env
docker compose up -d --build

Open http://localhost:7000 when the containers are healthy. The first admin password is printed in docker compose logs odysseus.

Building the Windows installer yourself

venv\Scripts\python.exe -m PyInstaller --noconfirm OdysseusFull.spec   # -> dist\OdysseusStandalone
& "$env:LOCALAPPDATA\Programs\Inno Setup 6\ISCC.exe" installer\odysseus.iss   # -> installer\Output\OdysseusSetup.exe

Native installs, GPU notes, macOS instructions, HTTPS, and configuration live in the setup guide.

Features

• Chat + Agents — local/API models, tools, MCP, files, shell, skills, and memory.
• Local backends — Ollama and llama.cpp both work as first-class targets. Odysseus asks each server what it can actually do (Ollama's /api/show, llama.cpp's /props) rather than guessing from the model name, so tool-capable models get native function calling instead of silently falling back to fenced text blocks. One context window setting in AI defaults governs both, and llama-server is launched to match it. If llamacpp_enabled is set, llama.cpp starts and stops with Odysseus.
• Remote execution (SSH) — set agent_execution_target to ssh:user@host and the agent's shell, python and file tools all run on that machine while the UI, database and model routing stay local. Aimed at work that is heavy to execute — Gradle, Flutter, test suites. Every tool routes through one execution environment (ported from Hermes Agent, MIT), so there is exactly one filesystem in play rather than files on one host and the shell on another. cd and export persist across tool calls via an env snapshot and an in-band cwd marker, and SSH connections are reused with ControlMaster. It does not move the model: an endpoint's base_url already accepts any host, so running a model too big for this machine is an endpoint change, not this setting. Round Table roles inherit it through their tool calls; the deterministic build gate still runs locally.
• Fast Mode — a chat-bar toggle that turns off a reasoning model's thinking step for quick turns. It sends the right signal per backend (reasoning_effort for Ollama, the chat template for llama.cpp), not just the /no_think prompt convention, so it works on both.
• Cookbook — hardware-aware model recommendations, downloads, and serving.
• Deep Research — multi-step web research with source reading and report generation.
• Compare — blind side-by-side model testing and synthesis.
• Documents — writing-first editor with AI edits, suggestions, Markdown, HTML, CSV, and syntax highlighting.
• Email — IMAP/SMTP inbox with triage, tags, summaries, and reply drafts. Agents can compose and send, gated behind an in-chat Approve / Decline card so nothing leaves without your click.
• Messages (macOS) — the agent can send and read iMessage/SMS. Sending is gated behind the same Approve / Decline card: just ask ("text Alex that I'm running late") and nothing leaves until you click approve. You can address people by name — contacts are resolved from CardDAV and your mail history, so "text Alex" works without you knowing the number. Delivery runs through Messages directly via AppleScript; no Shortcut setup is required. The first send raises a macOS Automation permission prompt for Messages — approve it once and it persists. Reading works too ("what did Alex say yesterday?"), straight from the local Messages database, read-only.
• Browser — the agent drives your real Chrome (your logins, bookmarks) over the DevTools Protocol on any platform. On macOS you can switch it to your real Safari instead — say "use Safari for the browser". Safari needs a one-time enable: run safaridriver --enable in Terminal, then Safari → Develop → Allow Remote Automation.
• Notes, Tasks + Calendar — reminders, todos, scheduled agent tasks, and CalDAV sync.
• Round Table — a SAFe multi-agent team (BSA → Architect → Developer → QA → Security → Tech Writer → RTE) that ships a ticket through real quality gates, with per-role model routing, a deterministic build gate, follow-up iterations, and human-in-the-loop merge. See docs/round-table.md.
• Themes — a borderless, minimal restyle with a dozen presets and live animated backgrounds, including a Matrix theme with digital rain.
• Extras — gallery/image editor, uploads, web search, presets, sessions, and 2FA.

Demo

A full hover-to-play tour lives on the landing page: docs/index.html.

Contributing

Help is welcome. The best entry points are fresh-install testing, provider setup bugs, mobile/editor polish, docs, and small focused refactors. See CONTRIBUTING.md and ROADMAP.md.

Security

Odysseus is a self-hosted workspace with powerful local tools. Keep auth enabled, keep private data out of Git, and do not expose raw model/service ports publicly. Deployment details are in the setup guide.

Star History

Stars for the upstream project this fork is built on (odysseus-dev/odysseus) — not this fork.

License

AGPL-3.0-or-later -- see LICENSE and ACKNOWLEDGMENTS.md.