Browser and code tools

Built-in tools lists every tool. This page explains how a few families work together, because a single tool does not tell you the sequence.

Browsing

Besides web_fetch and web_search, which read a page or a search result as text, the agent can drive a real browser with the browser_* tools. It is useful when a page needs a click, a form or some waiting before it shows what you want.

The sequence. The agent opens a page, takes a snapshot, then acts on what the snapshot shows:

1. browser_open opens a page, browser_navigate moves an existing one.
2. browser_snapshot captures a compact view of the page: its visible text and its interactive elements, each with an element_id.
3. browser_click and browser_type act on an element by its element_id, not by a CSS selector. The IDs come from the latest snapshot, so the agent takes a new snapshot after the page changes.
4. browser_press sends a key (Enter, Tab, Escape), browser_scroll scrolls, and browser_wait waits for the page to settle or for some text to appear.
5. browser_extract returns a readable text view of the page, with its headings, without the whole snapshot. browser_screenshot returns an image when the text is not enough.

Several pages. A session can hold several pages. browser_list_pages, browser_select_page and browser_close_page manage them. browser_search_content searches the text of the snapshots already taken, across pages.

Downloads and network. browser_list_downloads lists the files the page downloaded. browser_network_list shows recent requests, responses and failures, which helps when a page does not behave.

Making it lighter

browser_set_network_policy blocks requests for the whole session, and browser_get_network_policy shows the current rules. It takes a list of URL patterns to block, and a preset:

Preset	What it blocks
default	Only your own patterns
text_only	Images, fonts, audio and video
aggressive	The same, and the main analytics and tracking domains

Blocking what the agent does not read makes pages load faster and keeps its snapshots small.

Choosing the browser

In Go, ClientConfig.BrowserExecutablePath says which browser program to start, and ClientConfig.BrowserRemoteControlURL attaches the tools to a browser that is already running. See ClientConfig reference.

Understanding code: lsp

The lsp tool asks a language server about the code, which is more exact than searching for text. It needs a file path, and for most operations a line and a column:

Operation	What it answers
definition	Where a symbol is defined
references	Where it is used
hover	Its documentation and type
symbols	Everything defined in a file
implementations	The implementations of an interface or an abstract method
incoming_calls, outgoing_calls	Who calls a function, and what it calls
workspace_symbol	A search for symbols across the project

Language servers are set in the lsp section of the configuration file. In a project you have not trusted, that section is ignored.

Working in isolation: worktrees

enter_worktree creates an isolated git worktree in a folder named .worktree-... and switches the session into it, so the agent can try changes without touching your working copy. exit_worktree returns to the original folder. By default Seshat creates a branch for the worktree and deletes the worktree when you leave it.

Goals, plans and tasks

These tools keep the agent’s own work visible to you.

Tool	Use
create_goal, get_goal, update_goal	A persistent goal for the session, with its status and the usage so far
enter_plan_mode, submit_plan, exit_plan_mode	Agree on a plan before changing anything. See Runtime modes
task_create, task_update, task_list, task_get, task_stop	A checklist of the steps, and background tasks. ctrl+t shows it in the terminal interface

A turn does not end while the task list still has open items: the loop reminds the model, up to three times by default. See One runtime, one loop.