| name | pentagi |
| description | Autonomous penetration-testing AI agents. Registered as an external integration only; not activated by default. Useful reference for multi-step offensive workflows, sandboxing, and tool gating. |
| source | https://github.com/vxcontrol/pentagi |
| integration | external-agent |
| status | vendored-reference-offensive |
| vendored-at | third_party/pentagi |
| pinned-ref | main |
| fetch-requires | -AllowOffensive |
pentagi
Descriptor-only registration (P9 batch). The real project is a Go/TypeScript multi-agent pentest framework; integrating it requires explicit allowlist plus isolated network namespaces and is gated behind a future skills-install --allow-network-offensive flag.
Why registered
- Good reference implementation for multi-agent task decomposition under a strict tool-permission model.
- OctoCode's
RuntimeToolExecutor permission/circuit-breaker design overlaps in spirit.
Integration plan (deferred)
- Ship as MCP server adapter (
pentagi-mcp) with its own approval gate.
- Disabled by default; opt-in via workspace
.octocode/config.toml key integrations.pentagi.enabled = true.
- Never auto-triggered by general skill discovery.
Safety notes
Offensive tooling. Do NOT auto-invoke. All tool calls must require human approval.
