Agent Tools
A tool is a capability you grant an agent — a way for it to reach beyond the model and act on your data. When an agent decides a tool is needed, it fills in the tool's arguments, the tool runs server-side under the agent's permissions, and the result is handed back to the model to continue its answer.
Every tool call is org-scoped, permission-checked, PII-masked, traced, and costed. An agent can only use the tools you attach to it.
Tool types
Honeyframe ships the following built-in tool types, grouped by what they touch.
Data
| Tool | What it does |
|---|---|
| Look up records in a dataset | Look up or query rows from a dataset so the agent can retrieve a record of interest. |
| Append rows to a dataset | Let the agent write new rows into a dataset — you choose which columns it may write. |
| Run a SQL query | Run a SQL query against a connection and return the rows to the agent. |
Knowledge
| Tool | What it does |
|---|---|
| Search a Knowledge Bank | Retrieve relevant documents from a Knowledge Bank by vector similarity (RAG). |
| Search the agent's knowledge base | Retrieve from the calling agent's own attached knowledge base — auto-injected at runtime. |
LLM
| Tool | What it does |
|---|---|
| Query an LLM / Agent | Call an LLM or another agent with a purpose and system prompt, then return its answer. |
| Generate an image | Generate images from a text prompt via an image-generation model. |
Actions & integrations
| Tool | What it does |
|---|---|
| Call an HTTP endpoint | Call any external HTTP API. The agent's arguments become the request body or query string. |
| Send a message (Email / Slack / Teams) | Send a message through Email, Slack, or Microsoft Teams, with managed templates. |
| Call another agent | Delegate a question to another published agent (multi-agent / sub-agent). |
| Run Python code | Execute custom Python. Define an invoke(args) function that returns a string. Runs in an isolated sandbox with no credentials and capped CPU/time/memory. |
| Use a connector (built-in) | Expose a configured connector as a tool the agent can call directly. |
Defining a tool
Each tool has a name (a lowercase machine name the model sees), a description (what it's for — the model reads this to decide when to call it), and a typed set of parameters the agent must fill in. Write the description as an instruction to the agent: a clear, specific description is the single biggest lever on whether the model calls the tool at the right moment.
Attaching tools to an agent
Attach tools in the Agent Builder (for simple agents) or drop a Tool Call block into a structured visual agent. An agent only ever sees the tools you attach — there is no ambient tool access.
Governance
- Permissions — a tool runs under the calling user's org and project permissions; it cannot reach data the user couldn't.
- PII masking — masking policies apply to every tool result, including for anonymous webapp viewers.
- Tracing — every tool invocation appears in the agent's trace with its arguments, result, and latency.
- Cost — LLM-backed tools accrue tokens against the org's AI spend, visible in Administration → Monitoring.
See also
- Agents overview
- Connectors — the connections SQL and built-in tools run against