A hardened harness for running paperclipai/companies.sh safely.
Paperclip is an AI agent orchestration platform that lets you create and manage multiple "companies" — teams of AI agents (CEO, engineers, QA, etc.) that coordinate through a shared server. This repo wraps it in Docker isolation, network monitoring, and audit tooling so you can experiment without risk.
# 1. Copy and fill in scoped API keys (see security/PLAYBOOK.md)
cp .env.example .env
# 2. Start the sandbox (Docker, no network by default)
./scripts/start.sh
# 3. Add a company template
./scripts/add-company.sh paperclipai/companies/default
# 4. Open the dashboard (accept the self-signed cert warning)
open https://localhost:3100paperclip-sandbox/
├── docker/ # Container config (Dockerfile, compose, entrypoint)
├── config/ # Paperclip config overrides, network allowlist
├── data/ # (gitignored) Bind-mounted as PAPERCLIP_HOME
├── scripts/ # Safe wrappers: start, stop, add-company, monitor
├── security/ # Playbook, static scan, audit, analysis report
└── docs/ # Architecture reference, upgrade checklist
Everything runs inside Docker. The host never executes paperclipai code directly.
| Control | Detail |
|---|---|
| Filesystem | data/ bind-mount only — nothing writes to ~/.paperclip |
| Network | Disabled by default; allowlist mode for approved endpoints |
| Capabilities | --cap-drop ALL, --security-opt no-new-privileges |
| Memory | 1GB limit |
| Telemetry | Disabled (DO_NOT_TRACK=1) |
| Credentials | Scoped, rate-limited keys with spend caps |
| Monitoring | mitmproxy sidecar, post-run audit scripts |
Before running with real API keys, follow the incremental trust gates in security/PLAYBOOK.md:
- Static scan (no execution)
- Docker with mitmproxy allowlist (default mode)
- Normal operation with full network
- Ongoing version-pinned hygiene
Companies are created inside the running Paperclip server. This repo has no
company-specific code — all company data lives in data/ (the database).
# Add a pre-built company from the catalog
./scripts/add-company.sh paperclipai/companies/fullstack-forge
# Or create one interactively via the dashboard
open http://localhost:3100See the Security Analysis Report for a detailed risk assessment, static scan results, and mitigation plan. The full threat model and trust gate details are in security/PLAYBOOK.md.
A Getting Started guide is also available for first-time evaluators.