Get started Star on GitHub

An open-source, agentic network red teaming tool.

Agents that plan, act, observe and report inside a scope you define. Every action is checked by a scope engine outside the model, gated by the autonomy level you set, and written to an audit log. One static binary, your model, your machine.

Read the docs See the agent loop

Authorized testing only. A scope file is required and approval gates are on by default. Read the rules

~/engagements/acme-demoexample agent run, fictional lab
sample data, fictional lab. Agent decisions and counts are illustrative.
ws-0031 svc-backup files01 dc01 dc02 hop 1hop 2hop 3refused

One of three candidate paths in the sample run. The last hop is refused by the scope engine.

runs on
Linux amd64/arm64macOSWindowsDocker
fits into
GitHub ActionsGitLab CISARIF viewersNeo4j export(example list)
your model
local runtimes (Ollama, llama.cpp)OpenAI-compatible endpointsnone, deterministic planner(example list)
WHAT IT ISin short

One binary, one graph, one agent on a leash.

pwner discovers hosts and identities on a network you are cleared to test and builds a single graph from them. An agent, running on your own model or a deterministic fallback, proposes what to check next.

A scope engine outside the model decides what is allowed. You decide who approves. Every step lands in an audit log. It ships no exploits and no implants, and validation is read-only.

TOURdetails live on their own pages

The agent

An observe, plan, act, validate, report loop with three autonomy levels. The model's output is a request, never an action.

See the loop

Modules

Discovery, enumeration, path analysis and read-only validation. Each module declares what it touches.

Browse modules

Architecture

A small core, a scope engine and a graph store.

Read the design
RULESresponsible use
Authorized testing only. If you do not have written permission to test a network, do not point pwner at it.
Read the rules
scope file requiredNothing runs without one. Targets that are not allowed are denied by default.
approval gatesApprove-each-step is the default. The agent cannot edit its own gates.
append-only audit logEach action records the rule that permitted it.
QUICKSTARTexample commands

Try it on a lab network before anything else.

Install, write a scope file, check it, then run a playbook against a lab range. The full walkthrough is in the docs.

Open the docs Read the rules 
curl -fsSL https://pwner.example/install | sh

Examples, not live. The install URL and registry paths are placeholders.

quickstart.shexample
pwner init acme-demo          # writes scope.yaml to edit
pwner scope check            # fails closed on a bad file
pwner agent run playbook.yaml # lab range 10.0.0.0/24