Skills

A skill is a reusable capability you can attach to an agent. Skills live in the Library and are scoped to your active workspace. Some skills are bundled; others you author yourself.

Use skills when an agent needs to do something beyond plain text generation — call a function, run a script, query an external system, or wrap a Model Context Protocol tool.

Runtime types

Every skill has a runtime_type that determines how it executes.

Runtime	What it is	Where it runs
prompt	A prompt template the model expands	In-process
function	A built-in native function	In-process
python	A Python script	Bubblewrap subprocess sandbox
node	A Node.js script	Bubblewrap subprocess sandbox
shell	A shell command	Bubblewrap subprocess sandbox
wasm	A WebAssembly module	In-process via wasmtime
mcp	A tool exposed by an MCP server	The MCP server process

Capabilities

Each skill declares the capabilities it needs:

• Network access — whether the skill may reach the network.
• Filesystem access — whether the skill may read or write files.

Capabilities gate what a skill is allowed to touch. A skill that does not request a capability is denied it by the sandbox, even if its code attempts to use it.

The sandbox model

How a skill is isolated depends on its runtime:

• prompt and function run in-process. They are trusted native code with no separate sandbox.
• python, node, and shell run in a bubblewrap (bwrap) subprocess sandbox with namespace isolation. Network, filesystem, and environment access are restricted to the capabilities the skill declares.
• wasm runs in-process via wasmtime, isolated by the WebAssembly sandbox.

Caution

Subprocess and WASM skills only get the access their capabilities grant. Request the minimum a skill needs — broad network or filesystem access widens the trust you place in that code.

Publishing

A skill has an is_published flag. Publishing a skill makes it available for agents in the workspace to use. Keep a skill unpublished while you iterate on it, then publish when it is ready.

Creating a skill

Open the Skills tab in the Library. The skill list (SkillList) shows existing skills; the editor (SkillEditor) is where you author one.

1. In the Library, open the Skills tab and choose New skill.

2. Give the skill a name and description, then pick its runtime type (see the table above).

3. Provide the skill body — a prompt template, function reference, script, WASM module, or MCP tool binding, depending on the runtime.

4. Declare the capabilities the skill needs (network and/or filesystem). Leave them off unless the skill requires them.

5. Save, then test the skill. When you are confident it works, set is_published so agents can use it.

Tip

Skills that expose a callable interface appear in the Tools catalog, so agents can invoke them like any other tool.