Self-improving AI agent companies — a V2 fork of paperclipai/paperclip with persistent memory, MCP support, skills management, and a self-improvement loop.
Built by HarborSEO — an AI-powered SEO platform.
Paperclip is an open-source control plane for AI agent companies. This fork adds:
Agents resume the same Claude Code conversation across all tasks within a project — no more starting fresh every run. Comments on tasks, new assignments, and wakeups all continue the same session.
Agents accumulate knowledge that persists across sessions:
- Global memory — cross-project learnings, CEO/board feedback, preferences
- Project memory — codebase patterns, team preferences, prior decisions
- Memory is automatically injected into every agent run as system prompt context
- Sync from Claude Code — discovers MCP servers you've already configured in
~/.claude.json(one click) - Per-agent assignment — assign specific MCPs to specific agents (CEO gets Slack, DevOps gets AWS)
- Company defaults — set MCPs for all agents, with per-agent overrides
- At runtime,
--mcp-configis automatically passed to the Claude CLI
- User-installed skills from
~/.claude/skillsare now selectable per agent - Checkbox UI in agent settings — toggle skills on/off per agent
- Selected skills are symlinked into the agent's execution directory at runtime
- Post-run KPIs — every run records: completion status, tokens, cost, duration, errors
- Observations — CEO agent or board humans log patterns they notice over time
- Experiments — opt-in A/B testing of approaches per agent
- Analytics page — company-wide performance trends across all agents
Requirements: Node.js 20+, pnpm 9+
git clone https://github.com/IncomeStreamSurfer/paperclip-surfers.git
cd paperclip-surfers
pnpm install
pnpm devpnpm paperclipai rundocker build -t paperclip-surfers .
docker run -p 3100:3100 -v ~/.paperclip:/paperclip paperclip-surfers- Open the board → MCPs in sidebar
- Click Sync from Claude Code
- Your authenticated MCPs appear — check which to import
- Assign to company (all agents) or specific agents
- Open an agent → Memory tab
- Click Add Memory — choose global or project scope
- Memory is injected every time the agent runs
- Open an agent → Skills tab
- User-installed skills from
~/.claude/skillsappear with checkboxes - Check the skills this agent should have access to
- Open an agent → Performance tab
- See KPI trends, add observations, track experiments
This project is a fork of paperclipai/paperclip.
Original work © 2025 Paperclip AI — MIT License.
V2 additions © 2025 HarborSEO — MIT License.
All original code, architecture, and design belongs to the Paperclip AI team. V2 adds new capabilities on top without modifying the core orchestration engine.
MIT — see LICENSE