Skip to content

Native Tools — Overview

Tools let an agent take actions beyond writing a reply. A native tool is implemented by SquadOS itself: the agent receives its input/output contract and can call it when context and configuration allow. Some work without settings; others require a valid model, scope, or destination.

The agent must already be saved. During creation, Tools tells you to save first because every connection needs the agent ID.

To add, configure, or remove a native tool, the effective role needs agents.write on that agent. Organization resources can require additional permissions:

Resource connected to the agentWrite permissions
Native tool from the pickeragents.write
Knowledge bases and AutoLearn, managed under Knowledgeagents.write and bases.write
Composio, an organization HTTP tool, or an MCP serveragents.write and tools.write

A person with tools.view but without tools.write may still see an integration in the current picker and then receive an error when attaching it. Ask an administrator to review the role in that case; do not reconnect the account or recreate the tool.

  1. Open Agents and select a saved agent.
  2. Open Tools.
  3. Click Add Tool.

The picker builds different groups based on the resources available to the organization:

GroupContents
Organization ToolsHTTP tools already created in the organization
MCP ServersConnected servers and their discovered actions
Native ToolsSquadOS tools not yet enabled on this agent
Other ToolsComposio integrations activated under Tools in the administrative sidebar

The organization and MCP groups are hidden when no item is available. A connected native tool is also removed from the picker until it is deleted from the agent.

After the connection, its card returns to the Tools list. A green dot means Configured; a yellow dot means Incomplete configuration. Tools with a form open that form immediately after being added. The remaining tools are enabled in one click.

These are the current public options, sorted by their translated display name:

ToolWhat it lets the agent doSetting after adding
AutoPromptPropose or apply changes to its own instructionsChoose Propose and wait for approval — the default — or automatic application; access to organization context is a separate setting
Contact tagsRead, apply, remove, and create contact tags on demandNone
Conversation tagsRead, apply, and remove tags describing the conversation topicNone
Current Date and TimeRead the date and time in the organization’s effective timezoneThe panel shows the organization timezone and where to change it
Generate Image with LLMGenerate an image from textual instructionsSelect an image model; without one, the card stays yellow and the tool is omitted at runtime
HTML to ImageRender HTML and CSS as an imageDefault HTML, width, height, and format are optional
NotesRead and update structured contact- or conversation-scoped memoryNone
Transfer ConversationTransfer to another agent or human support based on the reasonCreate at least one context with guidance and a destination; schedules and conditions can restrict each route

Files also exists in the registry, but remains in preview and appears in the picker only for superadministrators. It needs a file- or tag-based scope; it can search the organization library and, when authorized, file generated content. The library does not have its own guide in this documentation yet.

Not every runtime tool is plugged in under Tools:

  • Search Knowledge Base is enabled under Knowledge when knowledge bases are attached to the agent.
  • AutoLearn is also controlled under Knowledge and logs opportunities for review when the agent cannot find a suitable answer.
  • Search earlier conversations is under Advanced, enabled by default. It can search only titles, topics, and summaries from the same person’s earlier one-to-one Hub conversations with this agent; it cannot read messages, rooms, direct person-to-person conversations, other users, or other agents.
  • Instagram private reply is made available automatically only while processing an Instagram comment whose inbox has Private reply via DM enabled. It is not shown in the agent picker.

Context blocks also live under Knowledge, but they are not tool calls: selected content is inserted directly into the agent’s instructions.

Organization tools are not native picker tools

Section titled “Organization tools are not native picker tools”
  • Composio connects third-party services. An account and its actions are first activated under Tools in the administrative sidebar; the integration can then be connected to the agent under Other Tools. See Organization tools.
  • Organization HTTP tools are reusable REST calls. They appear under Organization Tools and can also be created through New Tool when the person has access. See Custom HTTP Tools.
  • MCP servers expose a discovered set of actions. The picker opens another step where you choose which actions from that server to attach to the agent.

The legacy technical type http_api remains registered only for older HTTP tools. Do not add HTTP Call as if it were an independent native option: create or select the organization HTTP tool instead.

On each turn, the pipeline loads active connections and builds the contracts sent to the model. Some tools also have context requirements: for example, Contact tags works only when a contact exists, Search earlier conversations enters only a one-to-one Hub conversation, and Transfer Conversation offers only routes valid at that moment.

The model receives the technical name, description, and parameters — not a guarantee that it will call the tool. To make behavior more predictable:

  1. Keep only the tools the agent needs; every contract consumes context.
  2. Under Prompt, open the active-tools card and copy the suggestion containing {{tool:technical_name}} when available.
  3. State when to call the tool, what data it may send, and what to do after an error or empty result.
  4. Under Model, use a model that supports tools. Prompt diagnostics flag an incompatible model.
  5. Adjust Maximum tool calls under Advanced Settings when a task must chain several actions.
  6. Save and validate with Test Agent before enabling an external channel.

A green dot confirms only the configuration known to the interface, not a successful real call. Testing should cover valid input, an empty result, an external failure, and any persistent side effect.

  • Human Handoff and Transfer to agent are hidden for new agents. Their definitions remain registered only so legacy connections can still be opened and removed. Use Transfer Conversation for new configurations.
  • Send Multiple Messages became Split response into multiple messages under Advanced Settings.
  • Finish conversation and Analyze satisfaction were removed from both the editor and runtime; do not reference them in new prompts.