Concepts

Toolkits and skills

Toolkits and skills shape what an agent can do during a turn.

A toolkit is an action-selector grant set. It says which catalog actions an agent may use as tools.
A skill is a reusable instruction bundle that can request or narrow tools.
The tool manifest is the resolved, model-facing set of tools for one agent turn.

Toolkits

Toolkits grant action selectors. A selector can be exact, wildcard, or group-like, depending on the API request shape. Examples:

github.issue.add_labels
github.pull_request.*
slack.*
custom.*

Toolkits are not enough by themselves. The final tool surface is the intersection of toolkit grants and the agent principal's permissions. If a toolkit grants github.issue.add_labels but the agent lacks the matching actions.execute... permission, the manifest reports the grant as blocked.

Skills

Skills add reusable instructions and optional tool filters. Use them when a behavior should be shared across agents:

"Review Terraform safely."
"Write customer-facing release notes."
"Use the incident-response checklist."
"Format output as a Linear-ready issue."

Project skills can be created directly or imported from a Claude Code or Dive-style skill document.

Tool manifest

The manifest endpoint resolves the effective tool surface:

GET /v1/projects/{project}/agents/{id}/tool-manifest

Resolution considers:

Toolkits assigned to the agent.
The agent principal's role permissions.
Optional toolkit subset filters.
Optional per-invocation allowed_tools.
Optional active skill narrowing.
Catalog readiness for provider-backed actions.

The manifest returns callable tools, warnings, blocked grants, and action catalog metadata so the app and runtime can explain why a tool is or is not available.

Working with assignments

Agents own their assignment order:

mobius agents replace-toolkits agt_triager --file toolkits.json
mobius agents replace-skills agt_triager --file skills.json

Keep the set small. A narrow manifest reduces cost, prompt noise, and the chance the model chooses the wrong tool.

Common patterns

Pattern	Shape
GitHub triage agent	GitHub read actions, one issue-label write action, triage skill.
Slack support agent	Slack read/post actions, support-response skill, approval interaction before sending.
Data cleanup agent	Table read/write actions, one private memory table, no external provider writes.
Coding agent	Environment actions, repository actions, artifact publishing, narrow branch/PR tools.

Agents receive toolkit and skill assignments.
Actions are the catalog entries surfaced as tools.
Roles grant the action permissions that make toolkit entries callable.
Tables can be exposed to agents as durable memory tools.
Interactive reference shows the manifest contract.