Overview

Documentation Index

Fetch the complete documentation index at: https://tinytalk.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

Tools turn your agent from a question-answering assistant into one that can take action. When a visitor asks “what’s the status of order 12345?”, an agent with the right tool attached can call your order API, parse the response, and reply with the live answer — instead of saying “I don’t have that information.” The agent decides when to invoke a tool based on its description. You define the tool once; the AI model handles the orchestration.

​

How tools work

A tool is something your agent can call during a conversation to do work outside the chat. When the AI model decides a tool is relevant to the current message, it:

1. Picks the tool based on its name, description, and “when to use” hint
2. Collects the inputs by extracting them from the conversation or asking follow-up questions
3. Runs the tool through Tiny Talk’s secure runtime
4. Reads the result and weaves it into a natural reply

Tools are attached at the agent level — every conversation that agent handles has access to its enabled tools. You can disable a tool without deleting it; disabled tools will count against your plan limit.

​

Tool types

A Custom Tool points Tiny Talk at your own HTTPS endpoint. You define parameters, headers, and secrets, and the agent calls the endpoint whenever the conversation calls for it. Best when the action you want lives behind your own API. A Platform Tool is a first-party integration Tiny Talk maintains for you. You install it in a few clicks, configure what the agent should offer, and let the runtime handle the integration plumbing. Best when the third-party service is one we already support.

​

Available tools

​

Plan availability

The per agent limit caps how many tools a single agent can have enabled. Custom Tools and Platform Tools share this limit — for example, on Standard AI you could mix 1 Platform Tool and 9 Custom Tools.

​

Security and limits

Tiny Talk’s tool runtime is built to keep your network and your secrets safe. These guarantees apply to every tool type. Network safety

• HTTPS-only. The runtime refuses any non-HTTPS URL.
• IP literals (https://203.0.113.10/...) are rejected. Hostnames must be fully qualified domain names.
• Private, loopback, link-local, and cloud-metadata IP ranges are blocked. Tools cannot reach localhost, internal cluster addresses, or AWS/GCP metadata endpoints.
• Redirects (3xx responses) are not followed. The endpoint you configure is the endpoint the request hits.

Execution limits

• Request timeout — 10 seconds.
• Response size — 20 KB (responses are truncated past this).
• Steps per turn — Up to 8 model-tool round trips per visitor message. Most replies use 1–2.

Secrets handling

• Encrypted at rest with AES-256 in the same vault that stores other workspace secrets.
• Decrypted only at request time, in memory, scoped to the single tool invocation.
• Never returned to the browser — the dashboard sees only a masked preview.
• Logger redaction strips Authorization, X-API-Key, and similar headers from request logs.

Failure modes If a tool times out, returns an error status, or hits a security guard, the model receives a deterministic message it can react to (“Tool call timed out”, “Tool could not be reached”, “Tool returned status 404”). The agent can retry, fall back to another tool, or tell the visitor it couldn’t complete the action.

​

Credit cost

Tools do not consume extra credits. Credit cost is determined by the AI model selection. Reasoning models that decide to call tools don’t pay extra credits per tool. Network and execution limits (timeouts, response size) apply per tool call but don’t appear on your billing.

​

Roles and permissions

Tool management follows your workspace’s role-based access control: See Roles & Permissions for the full matrix.

​

Next steps