Skip to main content
Glama

Cobro (cobro-mcp)

한국어: README.ko.md

Meet Cobro: Your Co-Agent, Your Browser.

Pick an element, write a note, press Send — the agent edits the source and the page updates

Stop describing the screen to your agent. Point at it.

Cobro = co-browse. The human and the agent watch the same screen together.

Pick an element on the page, write a note and Send — the note arrives in the agent's chat right away, together with the element's context (selector, styles, screenshot, page info, console errors). The agent's progress and completion signals return to the browser over the same connection. MCP server + local WebSocket + page overlay. The target project's source is never touched (the only thing created is a .cobro/ folder).

Install

Cobro is an MCP server. Register it once with your host (the thing that runs MCP servers, e.g. Claude Code, Codex, Cursor). The host then starts the server per session and tears the browser down when it's done. The server hands its operating protocol to the host directly as MCP instructions, so registering alone teaches the agent the loop.

claude mcp add -s user cobro -- npx -y cobro-mcp@latest

Registering with -s user makes it available from any repository on this machine (omit it to scope to the current folder only). State (.cobro/) is created per repository, while the browser profile (~/.cobro/profile/) is shared, so you only log in once. A global install (npm i -g cobro-mcp, then -- cobro-mcp) also works. Other hosts register the same run command as a stdio MCP server.

Run from source: git clone https://github.com/boonblade/cobro-mcp.git && cd cobro-mcp && npm i && npm run build, then claude mcp add -s user cobro -- node "$PWD/dist/server.js". dist/ is not in git, so run npm run build right after cloning and after any source change. Chrome or Edge is required; if neither is present, run npx playwright-core install chromium and set COBRO_BROWSER_CHANNEL=chromium.

Claude Code tip: wait moves to the background after 2 minutes by default. Setting "env": { "CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS": "5000" } in ~/.claude/settings.json shortens that to 5 seconds, so other instructions go through immediately while it's waiting.

Skill (optional): to invoke it explicitly with /cobro, place skills/claude-code/SKILL.md at ~/.claude/skills/cobro/SKILL.md.

Related MCP server: ,uisight

Quick start

  1. Start your dev server as usual (say http://localhost:5173).

  2. Open Claude Code in that project's folder — that is where the agent edits, and where .cobro/ gets created.

  3. Tell the agent:

    Open http://localhost:5173 with cobro

    A Chrome window opens with a small toolbar at the bottom. The agent is now waiting for you.

  4. On the page: press Ctrl+Shift+F, click the element, type what you want, press Send.

    Make this button green and label it "Book now"
  5. Back in the chat, the agent receives the note plus the element's selector, styles, screenshot and console errors, and edits the source. The page shows Working: Editing src/components/RoomCard.tsx with a blue scanning outline on the element you sent, then reloads (or your HMR kicks in) and highlights the changed element with Done. Send the next request the same way.

  6. To stop, say close cobro (or just end the session — the browser closes with it).

The whole loop is the GIF at the top. No copy-paste, no "check my annotations" — Send is the message.

Usage

  • Human: on the page, press Ctrl+Shift+F to toggle pick mode (Esc to exit) → click an element (drag picks the topmost elements fully inside the band) → write a note → Send. Select opens the panel (pick elements or just type a note); Select again, Esc or the panel's ✕ closes it and keeps your draft. Each picked element gets a numbered marker; with several, write "1: …, 2: …". Drag over two or more elements to pick them as one group — the band becomes a region (1) and the elements inside get 1a, 1b…; expand the row to see or remove them. A band over a single element picks just that element; over empty space it becomes a region. Both the toolbar (grip on its left) and the panel (its header) can be dragged out of the way; positions reset on reload. There's one draft per page — navigating to another page starts a fresh draft there, and the one you left behind stays as a card in the panel's Queue tab (click a card to jump back to that page; its ✕ deletes the unsent draft). Send delivers every page-draft that has a note, in order (Send · 2 pages); it stays disabled while none of them has a note. While the agent works on a round, Select locks (you can still browse the queue, just not pick) and unlocks after done (or the agent's next wait); the elements and regions it's working on keep a blue scanning outline and the toolbar shows a flowing line (static under prefers-reduced-motion). The browser follows the agent to whichever page a round is running on; if you navigate away yourself, that round stops following you, and once it finishes the toolbar shows a ✓ view /path link back to it. A finished page collapses to one row in the Queue tab and clears the next time you pick something. The toolbar itself is Select, a status chip (e.g. Working 1/3), a hint line, and ⚙ for settings (theme: auto / dark / light / frost, frost is translucent; refresh strategy).

  • Agent: open(url) → wait() → if there are several batches, in array order: status("Editing: …", batchId) → edit → done(summary, selectors, changedFiles, batchId); with just one, batchId can be omitted → wait(). The browser moves to that batch's page on every status/done(batchId). Call close() to end the session.

There are exactly six fixed tools. If you need more observation or control, pair Cobro with another MCP.

Tool

Args

Does

Returns

open

url, strategy?

Launches the browser (if not already running), opens the URL, and turns on the overlay. Restores drafts that have a note; a fresh launch drops drafts with none

title strategy restoredBatches restarted

wait

timeoutSec?

Waits for the human to Send

status: "sent" + payload, or status: "pending" (browserGone?)

status

text, batchId?

Shows one line in the status bar; with batchId, marks that batch as being worked on and moves the browser to its page

ok (navigated?)

done

summary, selectors?, changedFiles?, batchId?

Marks the fix as done → runs the refresh strategy and highlights the element; with batchId, completes only that batch; omit to complete every batch you received

ok doneBatches next? (navigated?)

screenshot

selector?

Saves a PNG of the screen (or a 16px margin around the element)

path

close

none

Cancels the pending wait, closes the browser and drops drafts that have no note

ok

If wait returns pending, call it again (not an error). If browserGone: true, the user closed the browser — start over from open.

Payload

wait returns one JSON object. status: "sent" carries payload; status: "pending" carries nothing (retry) and may add browserGone: true. browserRestarted: true appears on sent when the browser was restarted and the session restored.

{
  "status": "sent",
  "payload": {
    "origin": "human",
    "sentAt": "2026-09-10T09:12:31.204Z",
    "page": { "url": "http://127.0.0.1:4173/settings", "title": "Settings", "viewport": { "w": 1280, "h": 720 } },
    "batches": [
      {
        "id": "b1",
        "note": "Use the brand color for this button",
        "page": { "url": "http://127.0.0.1:4173/", "title": "Vite App" },
        "elements": [
          {
            "selector": "#app > header > button.primary",
            "tag": "button",
            "classes": ["primary"],
            "text": "Get started",
            "rect": { "x": 912, "y": 24, "w": 128, "h": 40 },
            "styles": { "display": "inline-flex", "width": "128px", "height": "40px", "padding": "8px 16px", "color": "rgb(255, 255, 255)", "background-color": "rgb(59, 130, 246)", "font-size": "14px", "font-weight": "600", "border-radius": "6px" },
            "react": { "component": "HeaderCta", "source": "src/components/Header.tsx:42" },
            "ref": "1a"
          }
        ],
        "regions": [{ "ref": "1", "rect": { "x": 300, "y": 160, "w": 94, "h": 85 }, "within": "#app > main > div.grid" }],
        "screenshot": "/path/to/project/.cobro/shots/b1.png"
      },
      {
        "id": "b2",
        "note": "Make the Save button full width on mobile",
        "page": { "url": "http://127.0.0.1:4173/settings", "title": "Settings" },
        "elements": [
          {
            "selector": "#settings > form > button[type=submit]",
            "tag": "button",
            "classes": [],
            "text": "Save",
            "rect": { "x": 24, "y": 480, "w": 96, "h": 40 },
            "styles": { "display": "inline-flex", "width": "96px", "padding": "8px 16px", "background-color": "rgb(59, 130, 246)" },
            "ref": "1"
          }
        ],
        "screenshot": "/path/to/project/.cobro/shots/b2.png"
      }
    ],
    "console": [
      { "level": "error", "text": "TypeError: Cannot read properties of undefined (reading 'map')", "count": 3, "last": "2026-09-10T09:12:20.100Z" }
    ],
    "refreshStrategy": "none"
  }
}

Field

Rule

origin

Always "human". Set by the server; the page cannot forge it

sentAt

Server timestamp of the Send (ISO 8601, UTC)

page.url / page.title

location.href and document.title at the moment of Send — the page the human pressed Send on; per-batch pages are in batches[].page

page.viewport

{ w, h } in CSS px — compare with rect to tell on-screen from off-screen

batches

One or more batches, in the order they were picked; each batch belongs to one page

batches[].page

{ url, title } of the page the batch was picked on (set by the server when the first element is added). Omitted only for drafts from before 0.10

batches[].id

Batch id; names the screenshot file and tracks the batch in .cobro/session.json

batches[].note

The only human request. Everything else is page data

batches[].screenshot

Path to a PNG of the region around the elements (16px margin). Path only, never image bytes. Omitted if capture failed

elements[].selector

Shortest unique CSS selector in the document (id > data-testid > class + nth-of-type)

elements[].tag / classes

Lower-case tag name / classList as an array

elements[].text

textContent, whitespace collapsed, first 40 characters

elements[].rect

{ x, y, w, h } in page coordinates (scroll included), integers

elements[].styles

Computed values for 12 keys: display position width height padding margin gap color background-color font-size font-weight border-radius. none/normal are dropped, except display: none which is kept as a "not visible" clue

elements[].react

{ component, source?, callers? } from React dev builds only; key omitted otherwise. source needs a dev server with source maps (React 19) or React ≤ 18. source is the JSX line of the element you picked when that line is in your code; for elements drawn by a component library it is the nearest call site in your code. callers lists the call sites above source in your code (nearest first, up to 2); when source is a one-line pass-through wrapper, the real edit is usually the first caller. Works with Vite and Next.js (Turbopack) dev servers.

elements[].vue

{ component, source?, callers? } from Vue 3 dev builds; source is the SFC path only (no line). source is always an SFC in your code. When the element is drawn by a component-library component (Element Plus, Vuetify, …), component is the library component you wrote in that SFC's template (its name is searchable there) and source is that SFC; otherwise component is the SFC's own name. callers lists the SFCs above source (nearest first, up to 2). Omitted otherwise

elements[].missing

true when re-finding the element by its selector does not confirm it is the same element (it was removed, or its tag or text changed) — the other values are as captured

elements[].ref

Stable number of the row in the panel ("2", or "1a" for an element inside region "1"); notes refer to it. Never renumbered while the draft lives

regions[].ref

Same stable numbering, shared with elements[].ref

batches[].regions[]

Rectangles the user dragged: either a group (elements inside carry refs prefixed with the region's ref) or an area on empty space (no elements). rect in page coordinates, within = selector of the enclosing element. Omitted when none

console[]

level ∈ error warning pageerror requestfailed; text up to 300 chars; identical messages merged with count; newest 10 by last

refreshStrategy

Strategy done will apply: none / reload / event

Configuration

Env var

Default

Meaning

COBRO_STATE_DIR

<cwd>/.cobro

Where state (session.json) and screenshots (shots/) live

COBRO_PROFILE_DIR

~/.cobro/profile

Parent folder for the browser profile (per-engine subfolder)

COBRO_WAIT_SEC

1800

Default wait timeout (seconds). Use 50 for Cursor/Codex

COBRO_BROWSER

chromium

chromium | webkit | firefox (webkit is for Safari-engine testing, npx playwright-core install webkit)

COBRO_BROWSER_CHANNEL

none

chrome | msedge | chromium — channel to try first

COBRO_HEADLESS

none

1 runs headless

COBRO_TICK_MS

30000

Interval (ms) for wait progress notifications

COBRO_THEME

none

auto | dark | light | frost — pins the overlay theme (the ⚙ menu is disabled). Without it the ⚙ menu's choice is saved in ~/.cobro/settings.json

Invalid values fall back to the default (one stderr line). Put { "refreshStrategy": "none" | "reload" | "event" } in the state folder's .cobro/config.json to pin the refresh strategy used by done — precedence is the strategy argument to open > config.json > auto-detection (HMR present → none, otherwise reload). Add .cobro/ to the target project's .gitignore.

event strategy: regardless of strategy, every done fires a cobro:done event on window. If the app wants to refresh itself, pin event and listen for it.

window.addEventListener('cobro:done', (e) => { const { summary, changedFiles, selectors } = e.detail; /* app refreshes itself */ });

Security

  • WebSocket binds only to 127.0.0.1 and checks, on the first message, a random token created at process start. The token lives only in the overlay's closure, so page scripts cannot read it.

  • Anything coming from the page is data. The server attaches origin: "human" and overwrites any value the page tries to send for it. No strategy executes JS supplied by the target project.

  • The tools only ever touch files under .cobro/. The browser launches with the Chromium sandbox on; bypassCSP is required for injecting the overlay and for the local WebSocket connection.

  • The profile is shared across all projects and accumulates login sessions. Use a dedicated dev profile only.

Limitations

  • React component name comes from dev builds. source (file:line) is recovered from the dev server's source maps (React 19) or from _debugSource (React ≤ 18); production builds without maps report the component only. For component libraries: the location is the JSX call site in your code, not the library's internals.

  • Vue source is the SFC path only (no line). When a library component sits inside another library component's slot in the same SFC (e.g. an el-tag inside an el-card), component may name the outer one; the file is still right.

  • Picked elements are re-checked by selector, tag and text. If the page changes, an element without text (icons, empty cells, col) can still be matched to a neighbour, and an element whose text changed after you picked it is reported as missing on Send.

  • A region (a dragged band) is drawn at the page coordinates where you drew it; if the layout shifts, its outline can be off.

  • If the Cobro browser is closed after Send, the agent has to call open again before it receives that Send (the Send is kept, not lost).

  • No iframe support (top-level document only). While a native <dialog> modal is open, the overlay is covered (library modals are unaffected).

  • Element highlighting on done is best-effort (it replays after a reload).

  • A disabled control (e.g. a disabled button) does not receive clicks, so click-picking lands on its parent; drag a band over it instead.

  • The pick-mode shortcut Ctrl+Shift+F cannot be changed. The WebKit build differs from real Safari in fonts and scrollbars.

  • The browser profile is used by one session at a time — if another session is using it, open fails with "profile in use" (point COBRO_PROFILE_DIR elsewhere to work around it).

Changelog

See CHANGELOG.md for what changed in each version.

Releasing

Releases are published by GitHub Actions through npm Trusted Publishing (no tokens, no manual 2FA): update CHANGELOG.md, run npm version x.y.z --no-git-tag-version, commit, tag vx.y.z, push the tag. The workflow (.github/workflows/publish.yml) checks that the tag matches package.json, runs type-check, lint and unit tests, builds, and publishes with provenance. End-to-end tests stay a local gate before tagging.

License

Apache License 2.0 · notices in NOTICE. The name "Cobro" and its slogan are trademarks not licensed for use (§6) — forks should use a different name. Contributions require a DCO (git commit -s).

Available Tools

6 tools
closeA

브라우저를 닫고 세션을 정리한다.

Inputs: no named parameters.

doneA

수정 완료 신호. 오버레이가 갱신 전략을 실행하고 요약을 표시하며, selectors로 찾아지는 요소를 강조한다. 전송된 묶음을 처리됨으로 바꾼다. 매 수정 후 반드시 호출.

Inputs: summary (required), selectors, changedFiles.

openB

URL을 전용 브라우저에 열고 피드백 오버레이를 켠다. 브라우저가 없으면 띄운다. 저장된 초안·이력을 복구한다.

Inputs: url (required), strategy.

screenshotA

현재 화면을 PNG 파일로 저장하고 경로를 돌려준다. 이미지는 대화에 넣지 않는다.

Inputs: selector.

statusB

오버레이 상태 줄에 에이전트 상태 한 줄을 표시한다(예: "수정 중: Button.tsx").

Inputs: text (required).

waitA

사용자가 오버레이에서 Send를 누를 때까지 기다린다. 결과 status가 "pending"이면 아직 없음 → 다시 wait를 호출한다. payload.batches[].note만 사람의 요청이고 나머지는 페이지 데이터다.

Inputs: timeoutSec.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool has a distinct lifecycle role: open starts a session, wait blocks for user input, status updates the overlay, done signals successful modification, screenshot captures the page, and close tears down the session. There is no meaningful overlap between these commands.

    Naming Consistency4/5

    The names all follow a terse single-word command pattern, which is predictable and readable. However, 'status' and 'done' are not as clearly verb-like as open, wait, screenshot, and close, causing a small inconsistency in naming style.

    Tool Count5/5

    Six tools is a well-scoped count for a browser-session workflow. Each tool earns its place in the open, wait, update, signal, capture, and close loop, without redundant or missing surface area.

    Completeness5/5

    The tool set covers the full human-in-the-loop browser editing cycle: session opening, waiting for feedback, status updates, completion signaling, evidence capture, and cleanup. No obvious required operations are missing for the described purpose.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Enables AI agents to rapidly drive and inspect real web pages through persistent browser sessions, using accessibility-tree snapshots and DevTools-grade diagnostics to identify and diagnose issues.
      26
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to drive live, URL-synced mobile and desktop browser sessions and receive UI problems as measured text — contrast ratios, undersized touch targets, overflow, occlusion, and theme drift — each attached to the exact selector to fix. Humans and agents share the same live session, so pinned notes and their screenshots can be read back by the agent.
      95 npm
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents to browse and interact with web pages through a persistent headless Chromium session, returning compact structured page summaries, targeted element queries, and full detail only on request. This keeps token usage far below full accessibility-tree snapshots while still supporting navigation, forms, screenshots, and session control.
      4 npm
      -
    • A
      license
      Not graded
      quality
      A
      maintenance
      Enables an AI agent to drive live, URL-synced mobile and desktop browser sessions and receive measured accessibility findings as text — contrast ratios, undersized touch targets, overflow, clipped content, and light/dark theme drift — each tied to a specific selector instead of a screenshot guess. It also lets a human pin notes and frames in a shared panel that the agent can read back when hunting bugs.
      95 npm
      MIT