Agents (MCP)

Let a coding agent (Claude Code, Codex, Cursor and others) look at the site you're working on in Litr: the page, its console, its network calls, its errors and a screenshot. It reads what broke instead of you describing it.

Litr speaks the Model Context Protocol. Agents only ever see your developer tabs, pages served from this Mac (localhost, 127.0.0.1, .test, .local, .localhost), and they can only read, apart from reloading a tab.

Turning it on

In settings, set Agents (MCP) to On.

Tell your agent about Litr (once):

Claude Code: click Copy Setup for Claude Code in settings and paste it in a terminal, or type:
```
claude mcp add --scope user litr -- /Applications/Litr.app/Contents/MacOS/litr --mcp
```

Codex: add to ~/.codex/config.toml:

[mcp_servers.litr]
command = "/Applications/Litr.app/Contents/MacOS/litr"
args = ["--mcp"]

Others (Cursor, Claude Desktop, …) take the same command in their MCP settings:

{ "mcpServers": { "litr": { "command": "/Applications/Litr.app/Contents/MacOS/litr", "args": ["--mcp"] } } }

Reload your developer tab. Litr watches a page's console and network from its load on, so pages open before you turned Agents on show nothing there until reloaded.

Then ask, for example: "The Save button doesn't do anything on my local site. Check the console and network in Litr."

What an agent can see

Tool	What it gives
`list_tabs`	Your developer tabs: address, title, loaded or not, in front or not, error count.
`page_snapshot`	The page as an outline of its DOM: elements with their id, classes and key attributes, and their text. Scripts and styles are left out.
`query_selector`	The elements a CSS selector matches: HTML, text, position and size, and the computed styles that usually explain a layout bug (display, visibility, size, spacing, colours, fonts).
`console_logs`	What the page wrote to the console since it loaded.
`network_requests`	The page's `fetch` and `XMLHttpRequest` calls: method, address, status, time, content type and the start of text or JSON answers. And every other file it loaded, with timing and size.
`errors`	Script errors, files that didn't load and unhandled rejections, as in Developer tab….
`screenshot`	A picture of the tab as it is. The tab has to be on screen in Litr.
`reload`	Reloads the tab, for instance after the agent changed its code.

Without a tab given, a tool looks at the developer tab in front, else the one it looked at last.

While an agent is looking, the sidebar says so: An agent is looking, with the site. It goes away a minute after the last look.

What an agent can't do

See any other site: your mail, your bank, anything not served from this Mac. list_tabs doesn't even show them.
See forget tabs, cookies, saved passwords, or what's typed in a password field.
Click, type, open pages or run its own scripts in the page. That may come, once looking has proved useful.
Reach Litr from another Mac or another user: litr --mcp talks to Litr over a socket file only you can open, not a network port.

Page content is written by the page, and Litr tells agents to treat it as data, never as instructions.

Notes for developers

While Agents is on, Litr watches the console and network by wrapping console.log, console.warn and the like, fetch and XMLHttpRequest in developer tabs. In Web Inspector, console messages then point at Litr's wrapper rather than your line. Turn Agents off for the exact source location. Litr keeps the newest 500 console entries and 500 calls per page, and the first 4 KB of each text answer.