Tools and Skills

Tools give an agent capabilities beyond answering with text. A Skill is different in the import flow: it supplies instructions to create or propose changes to a prompt, but it does not install code or connect tools automatically.

Before you configure

Save the agent first. The Tools tab can only add resources after the agent exists. Editing the agent requires Edit agents (agents.write) in the applicable scope.

Organization resources also depend on Tools permissions. The picker can currently show integrations to a role without tools.write, with the error appearing only when that person tries to attach one. If this happens, ask an administrator for the appropriate permission; the limitation is recorded in the dossier as BUG-TOL-077.

What the Tools tab shows

The list combines the tools attached to the agent and sorts them by name. Each row shows the name, a status dot, and Edit and Remove actions.

• Green — Configured: the minimum configuration for that tool type is present.
• Yellow — Incomplete configuration: a required setting is missing. A Composio integration, for example, needs a selected account and at least one action.

The dot is not an operational test of the external service. Validate the tool in its editor and then use Test your agent before activating the agent.

Knowledge bases, AutoLearn, and past-conversation search are not managed in this list. Use the editor’s Knowledge and Advanced sections, as appropriate.

Add a tool

Select Add Tool. The picker can show four groups:

1. Organization Tools: HTTP endpoints already created for the organization.
2. MCP Servers: servers registered by the organization; after choosing one, select which discovered tools the agent may use.
3. Native Tools: capabilities built into SquadOS that are not yet attached to the agent.
4. Other Tools: Composio integrations that are active in the organization and not yet attached to the agent.

The HTTP and MCP groups only appear when a resource is available. If no Composio integration is available, Enable tools in Admin opens /admin/tools.

Administrators can also use New Tool without leaving the editor to create an HTTP endpoint or MCP server. The newly created resource then proceeds to the agent attachment flow.

What happens after selection

• Native: the tool is added. If it has settings, a configuration drawer opens; some tools with no settings are ready immediately.
• Organization HTTP: the attachment uses the central configuration. At agent level, you can fill only exposed parameter overrides and open Admin > Tools to edit the shared definition.
• MCP: a new attachment drawer starts with all discovered tools selected. Review the list and save; existing attachments preserve the previous selection and flag newly discovered tools.
• Composio: first select the allowed actions, then the connected account. Curation in Admin > Tools limits which actions can be selected for the agent.

See Native Tools — Overview, Custom HTTP Tools, and Organization Tools for the specific contracts.

How the agent chooses a tool

The model receives the technical name, description, and schema of the allowed arguments, together with the prompt and conversation history. Write descriptions that explain:

• the result the tool produces;
• when to use it and when not to;
• which data to ask for before the call;
• whether the operation reads or changes data;
• how to confirm a sensitive action with the user.

Connect only what is needed. Extra tools consume context and increase the chance that the model selects the wrong action.

Create an agent from a Skill

Create from Skill is on the Agents page, next to New Agent. The feature is in preview and, under the current configuration, is visible only to superadministrators who can also edit agents.

The importer accepts .md, .markdown, .skill, .txt, and a Skill folder compressed as .zip. The upload can be up to 1 MiB; the main text can be up to 100 KiB and 40 titled parts. In a ZIP, SquadOS looks for SKILL.md or another Markdown file, reads only compatible text resources, and does not execute programs or open binary files.

After you choose a file, SquadOS:

1. extracts its name, description, and titled parts;
2. attempts to adapt the procedure to known platform capabilities;
3. reports adjusted or removed steps, tools that may need to be enabled, and skipped files;
4. lets you choose between creating a new agent and suggesting changes to an existing agent.

Suggested tools are guidance: connect and configure each one manually.

New agent

The agent starts turned off, with the imported text as its prompt and the platform’s default model. Open the editor, read the complete instructions, connect the required tools, test, and only then activate it.

Existing agent

The file must contain titled parts (##). Each part becomes a proposal in the agent’s Suggestions tab. The prompt does not change until a person approves each proposal.

Skill security review

Treat every external Skill as untrusted content. It may contain instructions to ignore rules, disclose data, or misuse tools.

In the audited implementation, the preview shows the name, description, part titles, and adaptation report, but does not show the complete body that will be written. If adaptation is unavailable, the original text may still proceed to import. Therefore:

1. read the source file before confirming;
2. for a new agent, review the complete prompt in the editor before activation;
3. for an existing agent, read each suggestion diff before approval;
4. manually verify recommended tools — the adaptation inventory may not recognize registered HTTP and MCP resources;
5. if suggestion submission fails, open Suggestions before retrying because earlier parts may already have been created.

These limitations are recorded as BUG-SKL-107, BUG-SKL-108, and BUG-SKL-109 and were not fixed by the documentation.

Checklist before activation

• Remove instructions that request secrets, data export, or actions outside the agent’s purpose.
• Confirm tool names and schemas in the current editor.
• Restrict Composio actions and MCP tools to the minimum needed.
• Exercise read and write calls in a safe environment and verify the real effect.
• Check error messages, missing-data handling, and the need for human approval.
• Activate the agent only after reviewing the saved prompt and the complete test.