Tools — External Capabilities for Agents

A Tool extends an agent's reach beyond pure language generation. Tools let agents run shell commands, read and write files, control a browser, or call any MCP-compatible server — locally or over HTTP.

Scenario: A "Web Researcher" agent is given a Playwright MCP tool and a Filesystem tool. It navigates to URLs, extracts content, and saves summaries to disk — all without writing a single line of integration code.

Tool types

Creating a tool

Navigate to Tools → New Tool in the UI, or use the API. The form fields depend on the tool type:

To create a tool via API:

curl -X POST https://fleetq.169.58.89.204.sslip.io/api/v1/tools \
  -H "Authorization: Bearer YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "playwright-mcp",
    "description": "Browser automation via Playwright",
    "type": "mcp_stdio",
    "transport_config": {
      "command": "npx",
      "args": ["@playwright/mcp"],
      "env": {}
    },
    "status": "active"
  }'

Assigning tools to agents

Open an agent's detail page and use the Tools tab to attach tools. Each agent can have multiple tools with a priority ordering — lower numbers are preferred when the LLM must choose. Tools are stored in the agent_tool pivot table and resolved at execution time by ResolveAgentToolsAction.

MCP server discovery

FleetQ can auto-discover MCP servers running on the host machine or reachable via the network. Two MCP tools power this workflow:

• tool_discover_mcp — scans for available MCP servers and returns a list with their capabilities.
• tool_import_mcp — imports a discovered server as a new Tool record, ready to assign to agents.

Built-in tools

Built-in tools are provided by the platform and require no external server or configuration beyond enabling them.

SSH fingerprints

For tools that connect to remote servers over SSH, FleetQ supports fingerprint verification to prevent man-in-the-middle attacks. Retrieve known fingerprints via the tool_ssh_fingerprints MCP tool or the GET /api/v1/tools/ssh-fingerprints endpoint. Fingerprints are checked automatically when the tool initiates a connection.

Semantic tool selection (tool_search)

When a team has more than 15 active tools, loading all of them into every agent prompt wastes tokens and dilutes the LLM's focus. FleetQ solves this with semantic tool selection: tool descriptions are embedded with pgvector and the most relevant tools for the current task are retrieved per call.

• Embeddings live in the tool_registry_entries table (cosine distance, HNSW index).
• The tool_search MCP tool can be called directly by agents for explicit tool discovery.
• Retrieval threshold is ≥ 0.75 cosine similarity; tools below that floor are excluded from the agent's toolbelt for that turn.

Activepieces auto-sync (660+ integrations)

Point FleetQ at a self-hosted Activepieces instance and every one of its 660+ "pieces" (Stripe, Slack, GitHub, HubSpot, Salesforce, Notion, OpenAI, Google Sheets, …) becomes available in FleetQ automatically as mcp_http Tool records.

• Hourly sync job discovers new pieces and updates existing ones.
• Optional piece_filter limits which pieces are imported.
• SSRF-protected — the Activepieces URL is validated on every request.
• Trigger a manual refresh with the activepieces_sync action on the integration_manage MCP meta-tool.

Popular tools

FleetQ ships with 16 popular tools pre-seeded covering common integrations (Playwright, filesystem, git, databases, Slack, and more). All are disabled by default to avoid unintended access. Enable the ones you need from the Tools page — no configuration required for most of them.

MCP tools for tool management

The FleetQ MCP server exposes the following tools so agents and LLMs can manage tools programmatically:

API endpoints