Skills & plug-ins
Teach it once. It's yours forever.
Skills package your know-how into files the agent loads when they're relevant — paired with connectors and hard guardrails, they turn Raccoon into a plug-in platform for the way you work.
name: Inbox triage
description: Sort, label, and draft replies every morning
Loaded the moment a matching task starts
Drafts email. Can never run a command.
Know-how, access, and guardrails
A file format for know-how
A skill is a SKILL.md — instructions, bundled scripts, required connectors — that loads onto the agent's computer exactly when it's needed.
Three ways to add one
Describe it and AI builds it with you, upload an archive, or import any public repository — verified and custom skills behave identically.
Auto-activates when relevant
Skills wake up in matching modes and can be pulled in on demand — no re-explaining your process every session.
Hard guardrails, not vibes
Allow-lists and deny-lists per skill — build a plug-in that drafts email but can never run a command.
Pairs with your connectors
A skill brings the know-how, a connector brings the access — together they're a working plug-in for your stack.
Sends in subagents
Big jobs fan out to helper agents in their own sessions — each one inspectable, each sharing the same workspace.
Same open standard. Stronger machine.
Claude Code and Manus read the same open SKILL.md standard. Raccoon runs it with the most access — and the hardest guardrails.
Invoking
Raccoon AI: Auto-detected, pre-loaded
Claude Code: Manual /, or model's guess
Manus: Manual / command
Execution
Raccoon AI: Persistent cloud computer
Claude Code: Local machine (risky)
Manus: Cloud VM
Environment
Raccoon AI: Browser-native, feels local
Claude Code: Local IDE/CLI, technical
Manus: Web, terminal-first
Tools & connectors
Raccoon AI: Declared in the skill + modes
Claude Code: Juggle manually
Manus: Paired manually
Guardrails
Raccoon AI: Enforced allow + deny lists
Claude Code: Allow-list only
Manus: Not surfaced
Long tasks
Raccoon AI: Survives compression
Claude Code: Re-read after compaction
Manus: Re-read after compaction
Frequently asked questions
What exactly is a skill?
A SKILL.md file: a name, trigger conditions, required connectors, optional tools, guardrails, and any bundled files or scripts. When it's relevant, it loads onto the agent's computer.
How do I create one?
Three ways: describe it and the AI builds it with you, upload a skill archive, or import one from a public repository. You can read, edit, refresh, or remove any skill inline.
What makes this safer than pasting instructions?
Skills carry hard guardrails — per-skill tool allow-lists and deny-lists enforced by the platform, not by hoping the model complies. A skill that drafts email can be physically unable to run shell commands.
I already have skills from Claude, ChatGPT, or Manus — do they work here?
Yes. SKILL.md is an open standard read by 30+ tools — Claude Code, Codex, Gemini CLI, Cursor, Manus — and Raccoon speaks it natively. Import from a zip or a public repo and they run, gaining what the format alone doesn't give you: declared connectors with missing-connection warnings, platform-enforced guardrails, and auto-loading when the task matches.
What are subagents?
Helper agents the main agent can send out for parallel or focused work. Each runs in its own session you can open and inspect, and shares the same files.
Related use cases
More of the product
Connectors
Gmail, Drive, GitHub, and any MCP server — with per-tool permissions you control.
Workspace
Files, live app previews, a slides studio, and a backend console beside every chat.
Memory
Preferences and context that carry across sessions — transparent and editable.
AI Canvas
An image canvas with region edits, text fixes, upscaling, and video generation.
