Agentic Tools

Agentic tools are available in the interactive CLI chat session: run ragrails chat with no query argument. This is separate from the stateless rag.chat() SDK method, which does not execute tools. The SDK exposes the lower-level tool-calling API, but your application owns tool execution and approval.

api_call can make real HTTP requests. The interactive CLI validates tool calls first and asks for confirmation before side-effecting API calls.

Start the interactive chat

CLI

ragrails chat

Inside the REPL:

> /tools
> Look up the current docs page at https://example.com/docs and compare it to our stored answer.
> Call the staging API to create a test ticket with this payload: ...

Why this is CLI-only today

rag.chat() is a stateless RAG turn: it retrieves context, calls the LLM, and returns a ChatResult. It does not own a user session, approval UI, authorization policy, request allowlist, or audit log. The interactive CLI can safely provide built-in tools because it has a human in the loop. Before a side-effecting api_call runs, the CLI validates the arguments and asks the user to confirm the exact request. For SDK apps, those decisions are application-specific. Use Tool Calling to build a tool loop where your app controls validation, permissions, confirmation, execution, and logging.

Built-in tools

Tool	Confirmation	What it does
`web_fetch`	No	Fetches a URL and returns readable markdown text, trimmed to 8,000 characters.
`api_call`	Yes	Makes a live HTTP request with explicit `method`, `url`, `headers`, `params`, and `body`.

The /tools command lists registered tools, required fields, and confirmation policy.

Tool safety model

The model proposes a provider-neutral tool call.
Ragrails validates required arguments and schema-specific rules.
Tools requiring confirmation pause for user approval.
Declined or invalid calls are not executed; the model receives a tool result explaining why.
The loop stops when the model returns final text or reaches the maximum tool iterations.

api_call validation requires:

Requirement	Detail
Required keys	`method`, `url`, `headers`, `params`, and `body`.
Allowed methods	`GET`, `POST`, `PUT`, `PATCH`, `DELETE`.
Body requirement	`POST`, `PUT`, and `PATCH` require a non-empty `body`.
Headers/params/body	Each must be an object. Use `{}` only when genuinely empty.
Authorization	Placeholder values such as `<token>` are rejected.

When agentic tools help

Need	Tool
Check a live web page that was not indexed	`web_fetch`
Compare stored docs with a current external page	`web_fetch`
Fetch account details from an API after the user gives credentials	`api_call`
Create or update a resource through an HTTP API	`api_call`

Boundaries

Boundary	Current behavior
SDK `rag.chat()`	Does not call tools. Use Tool Calling and execute tools inside your own application boundary.
REST chat	Does not expose the CLI agentic tool loop.
Tool set	Built-in CLI tools are `web_fetch` and `api_call`.
Iterations	The tool loop stops after a bounded number of tool iterations.
Streaming	Interactive chat can stream plain answers, but `chat_stream()` does not stream agentic tool execution.

Reference: CLI chat and Tool Calling.

Get Started

Features

Capabilities

Use Cases

Walkthroughs

Start the interactive chat

Why this is CLI-only today

Built-in tools

Tool safety model

When agentic tools help

Boundaries

​Start the interactive chat

​Why this is CLI-only today

​Built-in tools

​Tool safety model

​When agentic tools help

​Boundaries

​Related pages

Start the interactive chat

Why this is CLI-only today

Built-in tools

Tool safety model

When agentic tools help

Boundaries

Related pages