XPI

Security tooling for the Pi agent: casefile tracking, web search, exploit-technique lookup, code search, and todos.

imagenpm version imagenpm downloads imageLicense: MIT

What it does

XPI adds a security workflow to the Pi agent: a SQLite case ledger, staged hunting with specialist agents (auditor, tracer, skeptic, chain), and a PoC gate that checks claims against real target/control behavior instead of trusting the model's say-so.

• Casefile: hypothesis → investigating → confirmed → reported, with gates at each step
• Attack transport: byte-exact raw_request, last-byte-sync race_send, multi-identity sessions, local JWT forging
• PoC gate: nonce-bound evidence, DNS-pinned target/control replay (or a same-host baseline for access-control bugs); OOB findings need an operator-run oracle; only the main agent can confirm
• Chains and primitives for multi-step findings, objectives for the kill chain
• Coverage matrix, plus AST search (ast-grep) and plain grep for code

Install

Works on Pi Agent and its fork OMP. One manifest serves both.

./install.sh            # auto-detects pi or omp in PATH
./install.sh --pi       # force Pi  (installs pi-subagents + optional ast-grep)
./install.sh --pi --no-subagents  # minimal Pi install; use /xp lite
./install.sh --omp      # force OMP (copies agents/*.md to ~/.omp/agent/agents)

Or from npm:

pi install npm:@xaccefy/pi-xpi     # Pi
omp install npm:@xaccefy/pi-xpi    # OMP

exploit_search needs PREVIEW_IS_API_KEY (see docs/guide.md).

Quick start

/xp        # toggle bounded swarm mode on/off
/xp lite   # single-agent workflow
/xp swarm  # bounded multi-agent pipeline

Swarm dispatches only auditor, tracer, skeptic, and chain. Validation, patching, reporting, and ConfirmFinding stay with the main agent.

Full tool reference and configuration: docs/guide.md.

Packages

Develop

bun install
bun test --isolate
bun run typecheck