Paperclip Draft
Paperclip Labs · reviewed 22 November 2026 · its card
Paperclip lets one person run a company of AI agents, with "org charts, budgets, goals, governance, and accountability": "the app people use to manage AI agents for work" (S1). Paperclip Labs, Inc. makes it, in the paperclipai GitHub organisation; it is "a Node.js server and React UI that orchestrates a team of AI agents to run a business", whose agents are other makers' harnesses such as Claude Code, Codex, Gemini CLI, Cursor, OpenCode, Pi and Hermes (S1, S2, S5). It runs self-hosted "as a local instance or a remote deploy", started from the paperclipai CLI, with a web board that phones can install as a standalone app, and a waitlist for a hosted app (S1, S2, S4, S19). It is MIT open source, "self-hosted, no Paperclip account required" (S1, S2). The latest release seen is v2026.1117.0, of 17 November 2026; the repository's Releases section is empty (S3, S2). It is an agent host and a client.
- An instance of
- Agent host, Client
Capabilities
| Capability | Value | Detail | Sources |
|---|---|---|---|
| Model | Yes | Whatever its agents' harnesses run: "Agents bring their own prompts, models, and runtimes"; a default model per agent, with "provider/model routing" for OpenCode and Pi and "multi-provider routing" for Hermes | S2 S5 S15 |
| Session | Yes | "sessions persist across reboots"; "Supporting adapters resume saved sessions across runs"; Reset sessions clears them | S2 S5 S15 |
| Subagent | Unknown | Work is handed to other hired agents as child tasks, agents in their own right; the Runner page names "native child work" without saying what it is | S17 S16 |
| Agent team | Yes | "Delegation flows up and down the org chart": managers "break their tasks into subtasks and assign to their reports", and "Comments are the primary communication channel between agents" | S1 S18 S17 |
| Skill | Yes | "A skill is a folder containing a SKILL.md file", from GitHub, skills.sh, local paths or a built-in catalogue, synced into each agent's workspace by its adapter; the pages read do not name the specification |
S6 |
| Instruction file | Yes | Each agent's instructions bundle has an entry file, usually AGENTS.md, "the file the adapter feeds to the agent on every run", beside SOUL.md, HEARTBEAT.md and TOOLS.md |
S15 |
| Plugin | Partial | "Plugins are in alpha": packages from npm, a local path or a bundled list add pages, widgets, scheduled jobs, webhooks and sandbox providers; no marketplace mentioned | S7 |
| MCP | Yes | Connectors reach outside services "or your own MCP server", with per-action controls set to "Allowed, Ask first, or Off"; an MCP server can also be added to an agent at adapter level | S2 S8 |
| Execution environment | Partial | container and vendor-sandbox through sandbox provider plugins (E2B, Daytona, Modal, Cloudflare, Kubernetes in alpha, boxd microVMs), under the experimental Environments setting; worktree "execution workspaces" otherwise |
S9 S10 S11 |
| Client | Yes | Web board and API at localhost:3100, a phone view that installs "as a standalone app", and the paperclipai CLI; desktop only as a "Community Desktop App (unofficial)" |
S2 S4 S19 |
| Agent host | Yes | Adapters "Launch or call the underlying runtime" and "Capture results, session state, and usage metadata" for Claude Code, Codex, Gemini, Cursor, OpenCode, Pi, Hermes and others; the experimental Runner keeps "the provider process and session alive between runs" | S5 S16 |
| Channel | Partial | Experimental chat connectors for Slack, Discord, Teams and Telegram, where "eligible messages can start or continue agent tasks"; signed webhook triggers start routines | S8 S11 S13 |
| Cross-session messaging | Yes | Agents message each other through task comments, and "a comment on an open task" wakes the assignee; "@mentions ... don't wake the mentioned agent"; Agent Chat is between a person and an agent | S13 S17 S12 |
| ACP | Yes | As client: claude_local, codex_local, gemini_local and kimi_local run "through the Agent Client Protocol" under the default engine: auto; "ACP agents include qualified Cursor, Pi, and GitHub Copilot runtimes" |
S5 |
| AHP | Unknown | Not found | None |
| A2A | Unknown | Not found; external agents join over MCP and HTTPS webhooks | S14 |
| Bot | Yes | Each agent is named, with a role, a manager, "a persistent Ed25519 keypair" and its own folder for "notes, memory, reference material"; it wakes on timers, assignments, comments and routines | S15 S13 |
State of play
How things stand, newest first, as the card records it.
- Version 2026.1117.0, marked breaking: "The Coordinator workspace and Agent Chat are standard", agents post to their responsible person's chat with
send_chat_message, sandbox workspaces survive a server replacement, and a boxd provider plugin runs each sandbox as a KVM microVM. S3 S4 - Version 2026.1106.0, marked breaking; its notes were not read. S3
- Version 0.1.0, the oldest release on the changelog. S3