Skills View

Browse installed skills and built-in tools per agent

The Skills view shows what tools and prompt skills each of your agents can use. It is a read-and-edit catalog: you can see every built-in tool, every platform tool, and every prompt skill — including which ones are allowed or denied per agent — and tweak the policy without leaving the browser.

Who it’s for: operators tuning what their agents can reach for, and authors checking whether a newly published prompt skill landed in the right agent’s workspace.

Route

• /skills — opens the Skills view

What You See

At the top of the page, an agent selector dropdown lets you pick which agent’s skill set you are inspecting. Below that, a scope toggle chooses between agent scope (only this agent) and shared scope (skills available to every agent). The rest of the view is organized into two tabs.

Built-in Tools tab

Lists Comis’s bundled tools, grouped by category:

Category	Tools
File Operations	read, write, edit, find, ls
Execution	exec, process
Search	grep
Web	webSearch, webFetch, browser

Below the bundled tools, a Platform Tools section lists higher-level building blocks the agent uses to talk to the rest of Comis — memory, sessions, agents, messaging, media, infrastructure, and fleet management groups (e.g., memory_search, sessions_spawn, discord_action, image_analyze, cron, gateway, models_manage).

Each tool card shows:

• The tool’s name and one-line description
• Notable parameter hints (for webSearch: freshness, deepFetch, provider override)
• An enable/disable toggle (controls whether this agent may invoke the tool)

A Tool Policy subsection lets you pick a profile (minimal, coding, messaging, supervisor, full) which sets the base allow-list, then add fine-grained allow/deny entries on top. Changes are persisted via the config.patch RPC and survive a restart.

Prompt Skills tab

Lists all discovered prompt skills for the selected agent, including:

• The skill name and description
• Its location (bundled, workspace, or local)
• Whether the skill is auto-injected or only invokable on demand

You can configure:

• Max body length — the largest skill body the LLM will ingest
• Enable dynamic context — whether matching context can auto-inject skill bodies
• Max auto-inject — how many skills can be auto-injected per turn
• Allowed / denied skills — pin lists for fine-grained control

Recent skill events (executions and rejections, fed by SSE) appear as a small log so you can watch which skills the agent actually reached for during a session.

Common Tasks

1. Switch which agent you are viewing

   Click the Agent dropdown at the top and pick another agent. The tool toggles, policy, and prompt skills all reload for that agent.

2. Disable a risky tool for one agent

   On the Built-in Tools tab, find the tool card (for example, exec). Click the toggle to off. The change writes to agents.<agentId>.skills.builtinTools.exec = false via config.patch and takes effect after the daemon restart it triggers.

3. Switch tool profile

   In the Tool Policy subsection, pick a profile from the dropdown (e.g., coding). The allow-list updates to the profile’s defaults; you can then add allow/deny entries for exceptions.

4. Allow only specific prompt skills

   On the Prompt Skills tab, add skill names to the Allowed Skills list. Anything not on the list is implicitly denied.

The Skills view is the closest visual analog to editing the agents.<agentId>.skills and agents.<agentId>.toolPolicy blocks in config.yaml. If you prefer the raw YAML, the Config Editor shows the same data.

Related Pages

Skills Overview

Concept guide for skills, tools, and the discovery layer.

Built-in Tools

Full reference for each bundled tool’s parameters and behavior.

Tool Policy

Profiles, allow/deny semantics, and security implications.

MCP Servers

Manage MCP-provided tools alongside built-in tools.