Odysseus_Code

A Claude-Code-Desktop–style coding agent for your self-hosted Odysseus.
One-click install · works with any API or 100% local · zero-risk to your existing setup.

---

⚠️ Alpha. The installer's zero-risk safeguards (backup → validate → auto-rollback) are implemented but not yet battle-tested across diverse Odysseus installs. Until this banner is removed, only run the installer against a backed-up or throwaway Odysseus, and prefer python install.py --dry-run first. Feedback welcome.

What it is

Odysseus_Code drops a Coding tab into an Odysseus install you already run — the same plan → search → write → verify agent loop you get in Claude Code Desktop, but on your models and your machine. Point it at an API (Anthropic, OpenAI, OpenRouter, Groq, …) or run it fully local.

It is an add-on, not a fork: it ships only the coding files + a safe installer and rides on your Odysseus core. Your existing Odysseus is never put at risk (see Zero-risk below).

Features

• 🧠 Agentic loop — plan, call tools, read results, verify, iterate, stop. Collapsed "Ran …" tool rows, live thinking, +/− diffs.
• 🔌 Works with API or local — Anthropic / OpenAI / OpenRouter / Groq, or Ollama locally. You choose at install time.
• 🗺️ Large effective context on a modest GPU — a deterministic, low-memory code-graph (symbols, callers, callees) + a session-recall store let the model reach a huge codebase precisely without holding it all in the window. (Recall, not a literal giant window.)
• 🛠️ Real tools — read/write/edit, bash, web search, repo map, git (status/diff/commit/PR).
• 🧩 Claude-compat — workspace .claude/skills/*/SKILL.md, .claude/commands/*.md, .claude/settings.json hooks (PreToolUse/PostToolUse), CLAUDE.md, and MCP servers.
• 🎨 Themed — matches Odysseus; Diff / Terminal / Files / Plan views; context wheel.

Install (one click)

Requires an existing Odysseus install.

1. Download this repo into your Odysseus folder (clone or download-zip):

   cd /path/to/odysseus
   git clone https://github.com/els-hub/Odysseus_Code.git

2. Run the installer (double-click or terminal):
   • Windows: double-click install.bat
   • macOS: double-click install.command
   • Any OS: python Odysseus_Code/install.py
3. The installer auto-detects Docker vs native, asks which model backend you want (API / local / both), installs only the missing dependencies, wires in the tab, and (Docker) rebuilds or (native) restarts. Open Odysseus → the Coding tab is there.

To remove it cleanly at any time: python Odysseus_Code/uninstall.py.

Zero-risk guarantee

The installer cannot break your Odysseus:

• It backs up app.py, index.html, and requirements.txt before any change.
• Every edit is idempotent and wrapped in odysseus_code sentinels.
• After editing it validates that app.py still parses; on any error it performs a full auto-rollback and leaves your install byte-for-byte as it was.
• If your Odysseus structure doesn't match what it expects, it aborts before touching anything.
• Run python install.py --dry-run to preview every change without writing.

Model setup

• API: the installer writes the provider key you give it into your Odysseus .env (see .env.example). Keys are never committed.
• Local: point at Ollama; the coding tab uses Ollama's native API so the context-window and reasoning controls work correctly.
• Both: pick the model per session from the model picker.

How the large effective context works (the honest version)

A modest consumer GPU can't hold a literal 100K-token window for a small local model. Instead the agent offloads its work to two deterministic, low-memory stores and pulls from them on demand: a tree-sitter/AST code-graph of the repo and a SQLite-FTS session-recall store of the model's own decisions/results. The live window stays small (fast, fully on the GPU); the model can reach far more than it holds — the same way Claude Code feels big on large repos.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. Good first issues are labeled.

License

MIT © els-hub