kimi-webbridge

Kimi Browser Extension (formerly Kimi WebBridge)

Control the user's real browser (with their login sessions) via a local daemon at http://127.0.0.1:10086 (the default address — see If a tool call fails for when it differs).

When the user asks to build a reusable CLI tool or script for a website (rather than do a one-off task), follow references/cli-creator/workflow.md. Its phases say when to read each of these:

• references/cli-creator/site-exploration.md — the protocol for finding how each feature gets its data
• references/cli-creator/login-handling.md — when the site requires login
• references/cli-creator/go-layout.md — when the CLI is written in Go
• references/cli-creator/companion-skill-template.md — the template for the skill that tells agents how to use the CLI

Tools

Tool	Args	Returns	Note
navigate	url, newTab(bool), group_title	{success, url, tabId}	First call opens a tab — see Tabs. group_title sets the group's visible label
find_tab	url, active(bool)	{success, url, tabId, borrowed}	Re-select a tab this session opened; active:true borrows the tab the user is viewing — see Tabs
snapshot	—	{url, title, tree} with @e refs	Accessibility tree (text) — use this to read page content and locate elements
click	selector (@e ref or CSS)	{success, tag, text}	Synthetic el.click()
fill	selector, value	{success, tag, mode}	Works on <input>/<textarea> AND [contenteditable] (ProseMirror/Lexical/Slate). mode is "value" or "contenteditable"
evaluate	code (for await, wrap it in an async IIFE — top-level await is a SyntaxError)	{type, value}
cdp	method, params	raw CDP response	Raw chrome.debugger passthrough — what evaluate is to JS, cdp is to CDP. Low-level escape hatch for cases the tools above don't cover. Page.bringToFront is a last resort: while focus emulation is off on a tab, a call turns it on instead of activating the tab (the tab renders and takes real input without being switched to); only a call once emulation is on activates, and it never pulls Chrome over another app the user is in. Target.activateTarget is refused
screenshot	format(png|jpeg), quality(0-100), optional selector (@e/CSS), optional path	{format, path, sizeBytes, mimeType}	Returns a file path, not base64 — see Screenshots
network	cmd(start|stop|list|detail), filter(URL substring, for list), requestId(for detail)	list: {count, requests:[{requestId, url, method, status, mimeType, completed}]}; detail: that request plus requestHeaders, requestBody, body (the response)	Capture is per tab, from start on. Call detail before stop — stopping releases the response bodies, and detail then returns bodyError instead of body
upload	selector, files(string[])	{success, fileCount}
save_as_pdf	paper_format, landscape, scale, print_background, optional path	{path, sizeBytes, mimeType, pageTitle}	Render current page → PDF, returns a file path — see Save as PDF
list_tabs	—	{success, tabs:[{tabId, url, title, active, groupTitle}]}	Inspect tabs in the current session
close_tab	—	{success, closed: bool}	Close the current tab in the session
close_session	—	{success, closed: int}	Close all tabs in the session — closed is the count. See Sessions for when to call

Tabs and the current tab

Single-tab tools (snapshot, click, fill, screenshot, save_as_pdf) act on the current tab — the one you most recently opened with navigate or selected with find_tab.

• Opening pages: use newTab:true when pages should coexist (comparing, cross-referencing); omit it to send the current tab to a new URL.
• Going back to an earlier tab: call find_tab to make a tab you opened earlier in this session the current one again. Pass the tab's URL — take it from list_tabs or the earlier navigate result. Matching is by host: kimi.com also finds a www.kimi.com tab, and the path is ignored. By default find_tab searches only this session's own tabs — it never reaches into the user's other tabs or windows.
• Acting on a page the user already has open: pass active:true ("use my open X tab" / "the X page I'm viewing"). It borrows the tab the user is currently viewing (returns borrowed:true); the borrowed tab is operated in place — it is not pulled into the session's tab group.
• If find_tab errors with "no tab matching … in this session", the page isn't open in this session — navigate with newTab:true instead.

curl -s -X POST http://127.0.0.1:10086/command \
  -d '{"action":"find_tab","args":{"url":"https://www.kimi.com","active":true},"session":"k26-research"}'

Call Format

Every command carries a top-level session naming the current task — see Sessions below. The examples in later sections omit it only for brevity; in real calls always include it. Every reply is an envelope: {"ok":true,"data":…} on success (the Returns column above describes data) or {"ok":false,"error":{"code","message"}} on failure. The command format depends on the user's OS.

macOS / Linux — inline JSON is fine:

curl -s -X POST http://127.0.0.1:10086/command \
  -H 'Content-Type: application/json' \
  -d '{"action":"navigate","args":{"url":"https://example.com","newTab":true,"group_title":"My task"},"session":"my-task"}'

If the inline call fails — a shell quoting/syntax error from bash, or an HTTP 400 from the daemon — do not retry the same command unchanged. Resend the request as a file body:

1. Write the JSON body to a uniquely-named temp file with your own file-write tool — never with shell echo/heredoc, which mangles the JSON the same way. Use a fresh name per request (e.g. /tmp/webbridge-req-<random>.json).
2. POST the file:

curl -s -X POST http://127.0.0.1:10086/command -H 'Content-Type: application/json' --data-binary @/tmp/webbridge-req-<random>.json

3. Delete the temp file as soon as the request returns.

Windows (PowerShell / cmd) — the shell corrupts non-ASCII characters (Chinese etc.) carried inline in command arguments or pipes; they reach the daemon as ? and the text is unrecoverable. Send every request as a file body instead:

1. Write the JSON body to a uniquely-named temp file with your own file-write tool — never with shell echo/heredoc, which corrupts non-ASCII the same way. Give every request its own filename with a random suffix (e.g. webbridge-req-<random>.json) so concurrent requests never share a file and overwrite each other.
2. POST the file with curl.exe — always curl.exe, never bare curl, which Windows PowerShell aliases to Invoke-WebRequest:

curl.exe -s -X POST http://127.0.0.1:10086/command -H "Content-Type: application/json" --data-binary "@$env:TEMP\webbridge-req-<random>.json"

3. Delete the temp file as soon as the request returns — don't leave request bodies on disk.

Sessions

One task = one session = one tab group. A session collects every tab the task opens into one tab group, so the user sees a single group for "what the agent is doing right now". Pass it as a top-level field of the request body (not inside args).

• Pick one session name at the task's start, put it on every command, and never switch mid-task — even across different sites. Switching session names per site is the #1 cause of fragmented tab groups.
• Name it after the task, not the site (camping-research, phone-compare). Use multiple sessions only for genuinely unrelated parallel tasks.
• group_title is the human-readable group label — write it in the user's language, on the first navigate of the task.
• When you create the group (the first navigate of a task), tell the user once that this task's pages are collected under group «title», and that you'll close them whenever they ask.

# First tab: set session + a human label (in the user's language)
curl -s -X POST http://127.0.0.1:10086/command \
  -d '{"action":"navigate","args":{"url":"https://www.kimi.com","newTab":true,"group_title":"K2.6 feature research"},"session":"k26-research"}'
# Another site, same task → same session → joins the same group automatically
curl -s -X POST http://127.0.0.1:10086/command \
  -d '{"action":"navigate","args":{"url":"https://www.moonshot.cn","newTab":true},"session":"k26-research"}'

Closing is always user-initiated: call close_session only when the user explicitly asks ("close those", "clear the tabs"). It clears the whole group in one call.

Screenshots

The daemon writes the image to disk and returns {format, path, sizeBytes, mimeType} — never base64, since the model can't read raw image bytes. Take the .path and open it with the Read tool to actually see it.

# Default: PNG of the visible viewport, daemon picks a temp path
curl ... -d '{"action":"screenshot","args":{}}'
# Options (each independent): JPEG quality, element-only via @e/CSS selector, custom output path
curl ... -d '{"action":"screenshot","args":{"format":"jpeg","quality":60}}'
curl ... -d '{"action":"screenshot","args":{"selector":"@e123"}}'

A caller-supplied path is honored verbatim (parent dirs created, existing file overwritten) — use a unique name to avoid clobbering. save_as_pdf follows the same rule.

Prefer snapshot over CSS/JS selectors

snapshot returns interactive elements with @e refs based on semantic role/name. Use them directly with click/fill — they survive CSS class hash changes that break manually-written selectors.

Fall back to evaluate (JS) only when:

• The target has no @e ref in the snapshot
• You need attributes not in the snapshot (e.g., href)
• You need to dispatch complex event sequences, or scroll

Evaluate Tips

• Always use compact JSON.stringify(data) — never add null, 2 formatting. Indentation and newlines can inflate the response several times over, causing truncation during transmission.
• evaluate calls share the page's JS realm — re-declaring the same const/let across two calls throws SyntaxError. Wrap in an IIFE for a fresh scope: (() => { const x = ...; return x; })().

Text input — use fill

fill (selector = CSS or @e ref, plus the value) works on <input>/<textarea> (returns mode: "value") and on [contenteditable] rich editors — ProseMirror, TipTap, Lexical, Slate, Quill, etc. (returns mode: "contenteditable"), firing the right input events so the page reacts.

fill is clear-and-insert: existing content is replaced. To append, read the current value via evaluate, concatenate, then fill with the result.

Form submit / special keys

There's no separate "press Enter" tool. To submit a form, click the submit button directly (click on the @e ref or selector). To dispatch a key event programmatically (e.g. Escape to close a modal):

{"action":"evaluate","args":{"code":"document.activeElement.dispatchEvent(new KeyboardEvent('keydown',{key:'Escape',bubbles:true}))"}}

Save the current page as PDF

save_as_pdf renders the current page to PDF and returns the file path. All args optional:

• paper_format: letter (default) | a4 | legal | a3 | tabloid
• landscape: false (default)
• scale: 1.0 (default), range [0.1, 2.0]
• print_background: true (default) — keep background colors
• path: caller-supplied output path; if absent, daemon picks a default under OS temp dir using the page title as the filename

path semantics match screenshot: written verbatim, parent dirs auto-created, existing files overwritten.

Decoded PDF cap is 100 MB. Above that the daemon refuses; reduce scale or split the page.

Known limitations

• Sites that strictly check event.isTrusted (some banking portals, captchas) ignore click / fill because those fire DOM-level synthetic events (isTrusted=false). For these, tell the user the page needs manual interaction. (Trusted input is possible at the protocol level via the cdp escape hatch, but treat that as advanced: your tabs are background tabs, and real input only reaches a background tab while focus emulation is on — send Emulation.setFocusEmulationEnabled {"enabled":true} before the input and {"enabled":false} when done with the tab. If input still does not land, Page.bringToFront is the last resort; see the cdp row.)
• Content inside iframes: snapshot, fill, click and evaluate see only the top document — an embedded document (preview pane, sandbox preview, embedded app) is a separate page and its nodes are not in the tree. When a snapshot lists frames after the tree (frames: [{frame:"#f1", url, size, …}], biggest first), read one with read_page {"frame":"#f1"}, using a handle from that list. A frame read echoes frame in its result; if it does not, the extension is older than this skill and returned the top page instead — see Version mismatches below. A preview frame often appears many seconds after the page itself loads — wait {"frame":"any"} returns the moment one filling a large part of the screen is readable, instead of re-snapshotting. A PDF preview is read through the viewer's own document, so every page comes back without scrolling. Acting on elements inside a frame is still not supported — navigating to the frame's URL directly sometimes works, but an authenticated preview stream usually loads nothing on its own.

If a tool call fails (daemon or extension not ready)

If a tool call can't reach the daemon (connection refused), or the reply isn't the ok envelope above (another program is answering on that port), run start yourself, don't ask the user. It is safe anytime: it no-ops if the daemon is already up. The binary is not on PATH: ~/.kimi-webbridge/bin/kimi-webbridge (Windows: %USERPROFILE%\.kimi-webbridge\bin\kimi-webbridge.exe).

Where the daemon listens is always in start's output, or in kimi-webbridge status (port when it is running, addr when it is not). If it isn't 127.0.0.1:10086, use that address for every call in this task. The config file is only for moving the daemon, never for finding it. Then retry the tool call.

If the reply is ok:false with code extension_not_connected, the daemon is fine and the browser side isn't — start won't change that. The message names what the user has to do; relay it in their language and carry on once they confirm.

start fails with "did not come up … held by another program" → another program owns the port. Move the daemon: write {"addr":"127.0.0.1:<port>"} (a free port of your choice) to the config file the error names, run start again, then ask the user to point the extension at the new port once (Kimi side panel → Settings → Local agent remote control → Connection address → ws://127.0.0.1:<port>/ws; it remembers).

If it still fails, the binary is missing (command not found means kimi-webbridge isn't installed), or the browser extension won't connect — point the user to the help page instead of deep-troubleshooting:

• English: https://www.kimi.com/en/products/kimi-webbridge
• 中文: https://www.kimi.com/products/kimi-webbridge

Never run stop / restart / upgrade / uninstall automatically — those kill a running daemon. When status shows version_mismatch, update_available or skill_mismatch, or a kimi-webbridge command prints a ==> … Run: <command> line (the same notices, on stderr), tell the user and hand them the command verbatim to run — the command field, or what follows Run:; that is what it is for. When the user asks about kimi-webbridge itself (install, start/stop/upgrade, "is it running", what status means, moving it to another port), read references/operations.md first.

Version mismatches

If a tool returns an error containing "Please update the Kimi Browser Extension", the user's browser extension is older than this skill. Don't try to reconcile versions yourself — just tell the user, in their language, to update the extension and retry:

• English: https://www.kimi.com/en/products/kimi-webbridge
• 中文: https://www.kimi.com/products/kimi-webbridge