A command-line agent for coding and cybersecurity tasks that runs against a local model (llama.cpp) via Goose. You describe a task in natural language; Legate classifies it as coding or security, mounts the relevant tools, and runs it.

Legate is a Python controller around Goose and llama.cpp. On top of Goose it adds:

• intent and kill-chain-phase routing that narrows a large MCP tool surface (e.g. Praetor, ~390 tools) to a working set a small model can handle;
• an optional separate (uncensored) model slot for authorized offensive work;
• hardware detection and model-fit recommendation for local serving.

• Python 3.11+ and uv
• git
• Goose CLI (goose)
• llama.cpp llama-server, build b10435 or newer (older builds drop Qwen3-Coder tool calls)
• An NVIDIA GPU for serving (CPU-only serving is not supported)
• Optional, for security work: an MCP server such as Praetor and its backends (Burp Suite, CLIs like nmap / katana / gau)

bash scripts/setup.sh

Installs Python dependencies, creates .env from .env.example, and auto-detects llama-server, the model file, and Praetor. It fills only blank values. To configure manually: cp .env.example .env and edit the LEGATE_* variables.

Serve a model, then run the agent:

bash scripts/serve.sh &

legate                                           # interactive session
legate chat "fix the failing test in ./parser"   # one-shot coding task
legate chat "recon http://target/ and list URLs and parameters"
legate chat "test the login at http://target/ for sql injection"

• Coding requests use the developer editor (edit / write / shell).
• Security requests are routed to a kill-chain phase (recon / scan / exploit / post / report) and driven through Praetor, plus the shell.
• In an interactive session, context carries across turns and output streams as it is produced.

Live security engagements require Praetor configured (LEGATE_PRAETOR_* in .env) and its backends running. Without them the Praetor tools error and the agent falls back to shell tooling.

Additional subcommands:

legate doctor              # report detected hardware
legate models recommend    # rank local models by fit for this machine

• Routing. A request is classified (coding vs. security, and for security which kill-chain phase) deterministically before the model runs. Only the matching phase's tools, ranked by BM25, are mounted — a small model cannot be given hundreds of MCP tools at once.
• Execution. Goose runs headless with GOOSE_MODE=auto in the working directory you specify. Goose owns the edit/write/shell tools; scope the agent to a directory you are willing to let it modify. Legate does not sandbox Goose's tool calls.
• Streaming. Goose's stream-json output is rendered live (assistant text, ⚙ tool(args), ↳ result). Interactive context is kept via Goose sessions; a change of phase starts a new session so the new phase's tools are mounted.
• Timeouts. Each model run is time-bounded — the Goose process is killed past the deadline, so a runaway generation or a dead model server fails fast.

Configuration is read from .env (see .env.example).

uv run pytest

Apache-2.0. See LICENSE.