What Is a Capability?
A capability is a self-contained folder inbrain/cerebellum/ that gives Wolffish a new ability. Capabilities are how Wolffish learns to do things — run shell commands, read files, interact with git, search the web, send messages.
Two Types of Capabilities
Pure Skills
A SKILL.md file only. The markdown body contains instructions the LLM follows using its existing abilities. No code needed. Example:
git/ — the LLM uses shell commands to run git operations based on the SKILL.md instructions.Plugin Capabilities
A SKILL.md file plus a
plugin/ folder with executable code. The plugin exports tools that the LLM can call directly. Example: shell/ — the plugin wraps child_process to provide a shell_exec tool.Capability Folder Structure
How Capabilities Are Loaded
On startup,cerebellum.ts scans brain/cerebellum/ and for each folder:
- Reads and parses the SKILL.md frontmatter (YAML)
- Extracts tool definitions, danger patterns, confirm patterns, and triggers
- If a
plugin/folder exists, dynamic-importsplugin/index.mjs - Calls
init(context)on the plugin, passingpluginDirandworkspaceRoot - Registers the tools in the cerebellum’s registry — every capability becomes visible in the capability index, but only a subset of tool schemas ships to the model (see below)
How Capabilities Reach the Model
Registering a capability is not the same as sending its tools to the model. Wolffish separates knowing a tool exists from carrying its schema — that’s what keeps the system prompt lean no matter how much you install.The capability index
Every system prompt contains a<capabilities> index: one line per capability — name, a short description (capped at 90 characters), and its tool count, with a [loaded] marker on the ones whose tools are callable right now. The whole index costs ~800 tokens, and past 60 capabilities the unloaded remainder collapses to a grouped count, so prompt cost stays flat in the number of installed capabilities. The model always knows what exists; it just doesn’t pay for schemas it isn’t using.
Core capabilities — always loaded
A bootstrap set ships its full tool schemas with every request (~31–49 tools):tool-discovery, introspect, filesystem, shell, ask, utilities, web-search, secrets, system, the operating-manual, pdf-design, dataviz and web-design manuals (one schema line each — their manual bodies load only when called), video (a render request must reach the tool without a discovery hop), plus workflow (gated to the workflow master — agents never see its tools) and the channel capabilities (telegram, whatsapp, electron, and phone for notifications) while connected. These are the keys to everything else — discovery, memory retrieval, files, and replying — and must never require a discovery hop.
Everything else — discoverable on demand
GitHub, Google, Notion, the browsers, computer-use, the media and document tools, and every MCP server load when needed:tool_search(query)— term search over capability names, descriptions, triggers, and tool names. The best match is auto-loaded and its tools are callable the same turn.tool_activate(capability)— explicit load by exact name, from the index or a search result.
github in a heartbeat run never invalidates a live chat’s prompt cache. At most 10 non-core capabilities stay loaded per conversation, evicted least-recently-used — and eviction is invisible, because calling an evicted tool auto-reactivates its capability and the call executes normally. The same courtesy applies to any known-but-unloaded tool called directly: it just works. Only a genuinely unknown tool name returns an error, and that error points at tool_search.
To make extra capabilities always-loaded, add their names to pinnedCapabilities (a string array) in config.json — there’s no UI for it; hand-edit the file.
Pure skills
Pure skills (no tools) appear in the index as(guide) entries. tool_search finds them by name, description, or trigger keywords, and the model reads their instructions on demand with skill_read_source. There is no keyword-triggered injection of skill bodies into the prompt — the index line is all a skill costs until it’s actually needed.
Built-In Capabilities
Wolffish ships with these capabilities in the default workspace:Adding and Removing Capabilities
You can add your own capabilities two ways:- Drop them onto Settings → Capabilities — a single
SKILL.md, a folder, or a.zip. Each drop is validated before anything is written, then loaded immediately. Imported capabilities show an Unknown badge (they’re yours, not bundled) and can be removed with the trash icon, which cleanly deletes their folder. - Create the folder by hand under
brain/cerebellum/, then click Resync in the panel (or restart Wolffish) to load it.
When a settings page depends on a capability
Most service pages in Settings are the front end of a capability — Computer Use, Memes, Speech-to-Text, Text-to-Speech, GitHub, Notion, Brave, Video generation. Turn that capability off (or have it fail to load) and the page used to sit there with controls that quietly did nothing. Each panel now checks for itself and says which of three things is wrong — the capability is missing, switched off, or failed to load (showing the underlying error) — with a button straight through to the Capabilities page. The controls underneath go visibly inert rather than pretending to work.Create Your Own
Learn how to build a new capability from scratch.