A coding agent that runs entirely on your hardware. Point Thoth at Ollama, llama.cpp, or any OpenAI-compatible server and it explores your codebase, edits files, runs builds and tests, and searches the web when it needs documentation. Every action is shown on screen, destructive ones wait for your approval, and nothing ever leaves your machine.
- Local first. Auto-detects Ollama and uses its native API to control the context window itself, so agentic prompts don't get silently truncated. Any other OpenAI-compatible server works too.
- Safe by construction. The model cannot edit a file it hasn't read, or overwrite one it hasn't read completely. These rules live in Rust, not in the prompt. Full diffs and full command lines are always displayed, even after "always allow".
- Editor aware. With the VS Code extension, Thoth knows your active file, selected lines and the Problems panel, and the model can re-check diagnostics after every edit.
- Built for small context windows. Live token meter, automatic compaction, per-project memory and session recaps keep long tasks running on 16-32k tokens.
- Self-sufficient. Web search through DuckDuckGo with no API key, or Google Programmable Search if you have one.
Linux / macOS:
curl -fsSL https://raw.githubusercontent.com/thoth-coder/thoth/main/scripts/install.sh | shWindows (PowerShell):
irm https://raw.githubusercontent.com/thoth-coder/thoth/main/scripts/install.ps1 | iexOr build from source: cargo build --release (-> target/release/thoth).
Update later with thoth upgrade.
ollama pull qwen3:8b # any tool-calling model
thoth # or: thoth -p "one-shot prompt"| Getting started | Ollama, llama.cpp, choosing a model |
| Usage | keys, commands, memory, editor integration |
| Tools & guardrails | what the model can do and how it is contained |
| Configuration | config file, env vars, context window |
Bug reports and pull requests are welcome. See CONTRIBUTING.md for the ground rules and SECURITY.md for the threat model.
Dual-licensed under Apache-2.0 or MIT, at your option.