Extending Imaginne

You can broaden what the agent can do and constrain how it acts. This section pulls together the extension points for power users and administrators.

The extension points

🧰Skills

The primary way to give the agent capabilities — auditable packages, governed by your organization.

Open →🔌MCP servers

Connect external tools via the Model Context Protocol. They appear to the agent as native tools.

Open →🧑‍🤝‍🧑Subagents

Focused specialists that the agent (or you) spins up in parallel for well-scoped tasks.

Open →🚧Workspace rules

Define, in a file, what the agent may read/write/run and what requires confirmation.

Open →

How to choose

You want to…	Use
Teach the agent a new task/process	Skills
Give the agent access to an external system/tool	MCP servers
Parallelize exploration, analysis, implementation	Subagents
Limit what the agent can touch in the project	Workspace rules
Control how much it asks before acting	Autonomy

Capabilities vs. guardrails

Think in two directions:

• Capabilities (what the agent can do) come from skills, MCP servers, and subagents.
• Guardrails (what it may do and when it asks) come from autonomy, workspace rules, and organization policy.

What about hooks, custom commands, or plugins?

To set expectations clearly, some mechanisms common in other tools are not yet user-configurable in Imaginne:

• Script-based lifecycle hooks (running a command on "before/after a tool," "session start," etc.) are not exposed today. The equivalent control — blocking, allowing, or requiring confirmation for actions — is handled through workspace rules and autonomy.
• Custom slash commands (/something) and functions can't be defined by users — the set is fixed. See the TUI reference.
• There is no agent plugin system beyond skills and MCP. (The Desktop plugins — advanced Preview, Mermaid, CAD, Canvas — are interface viewers, not agent extensions.)

The real, supported extensibility is: skills + MCP + subagents, with workspace rules and autonomy as guardrails.

Start with skills

For most cases, a skill is the way to go — auditable and governable by your organization. Use MCP when you need to talk to an external system, and subagents when you want to parallelize.