Revenant

A local, offline coding-agent CLI powered by your own Ollama models — a Claude-Code-style tool-calling loop that runs entirely on your machine. No cloud, no telemetry, no API keys.

pip install revenant-cli

This installs the revenant command.

Usage

# investigate (read-only)
revenant "summarize what the auth module does"

# make changes (each mutating action asks for approval)
revenant "add a /health route to the server"

# point at another repo, pick a model
revenant --workspace ~/proj --model qwen2.5-coder:7b "where is rate limiting handled?"

Revenant reads and searches your code (read_file, glob, grep), and — with your approval — edits files and runs shell commands (write_file, edit_file, run_bash). Destructive commands (rm -rf /, fork bombs, …) are hard-blocked even in --yolo mode.

Requirements

• Python 3.11+
• Ollama running with a model pulled (e.g. ollama pull qwen2.5-coder:7b). Revenant talks to a local model server; it does not include an LLM.

A tool-capable model (the qwen2.5 family) is recommended — Revenant auto-detects native tool-calling and falls back to a prompt-based protocol for models without a tool template.

Flags

--workspace (repo root, confined) · --model · --max-steps · --max-context-tokens · --read-only · --yolo (auto-approve) · --no-native-tools.

License

MIT. Part of the Revenant monorepo. Built on nerva-core (local-LLM layer) and nerva-agent (the agent engine).