MCP servers › io.github.Gwakko › agent-in-browser
MCP server · registry snapshot 2026-10-23 · last tested 2026-10-23
io.github.Gwakko/agent-in-browser
Real browser for any agent: Chrome/Edge/Brave via CDP, Firefox via BiDi. Refs, forms, console, net. — the publisher's description
Summary
io.github.Gwakko/agent-in-browser is a local MCP server listed in the official MCP registry at version 0.1.1. On 2026-10-23 ProtocolProbe started its published package in an isolated sandbox in read-only mode and read its inventory: 23 tools, 0 resources and 0 prompts.
Record checks
4 of 5 record checks met · 4 do not apply
- met: Registry entry is active and says how to run it
- met: Source repository link opens
- met: Repository is not archived
- met: Code pushed to the repository in the 90 days before our check
- not met: Repository publishes a security policy (SECURITY.md)
- not applicable: Package names its licence with a standard (SPDX) identifier
- not applicable: The version the registry lists can be installed and is not withdrawn
- not applicable: npm package carries build provenance (attestation)
- not applicable: No known critical or high advisory in the package or its dependencies
- Not ranked by downloads: not an npm or PyPI package
- Not ranked by stars: 0 stars (ranked from 10)
From public records, not a safety rating. Each check with its reason · Is it safe to install?
Key facts
- Runs
- local package
- Transport
- standard input/output
- Package
- MCP bundle
- Repository
- public
- Tools
- 23
- Described as changing data
- 16 of 23 toolsAI-classified
- Mainly works on
- a web browser or desktop applicationsAI-classified
- Topic
- search-and-web-dataAI-classified
- Last tested
- 2026-10-23
Worth knowing before you connect it
- 16 tools are described as changing something (browser_close, click, console_messages, …): our AI model's reading of the descriptions, not something we tested. We did not call them.
Each line comes from a section below, with its source and date.
How fresh this is
- Live-tested
- yes, a real session (not a text-only review)
- Last tested
- 2026-10-23
- Mode
- read-only (no tool calls)
- Latest version
- 0.1.1, published 2026-10-21 (registry, as of the 2026-10-23 snapshot)
- First listed version
- 2026-10-21 (earliest version still in the registry; older removed versions would be earlier)
- Registry data
- snapshot 2026-10-23
- Package and repository
- checked 2026-10-23
- Engine
- mcp-runtime-engine@0.1.0 · sdk@1.32.0
Changes since the last refresh
New in the registry: not in the list of 2026-10-04, listed on 2026-10-23 with version 0.1.1 (published 2026-10-21).
Source: the official MCP Registry list on 2026-10-04 and on 2026-10-23.
What it is
What does this server offer?
Observed inventoryWhat it mainly works on, as our AI model read its tool descriptions: a web browser or desktop applications.
16 of 23 tools are described as changing something (creating, editing, deleting, sending or running).
“Close the browser launched by agent-in-browser.”
Publisher declares: read-only no · open-world yes
Takes no inputs.
“Show the connected browser (launching it if needed), protocol, current tab and open tabs.”
Publisher declares: read-only yes · open-world yes
Takes no inputs.
- browsers_installedreads or searches
“List browsers detected on this machine and whether they can be driven.”
Publisher declares: read-only yes · open-world yes
Takes no inputs.
“Click an element (by ref, selector or visible text) or page coordinates with real mouse events.”
Publisher declares: read-only no · open-world yes
8 inputs
buttonstring“Mouse button (default left).”doubleboolean“Double-click.”matchstring“Visible text / accessible name to match (case-insensitive substring).”refstring“Element ref from snapshot/find, e.g.”selectorstring“CSS selector (first match).”tabstring“Tab id from tabs_list.”xnumber“Viewport x coordinate (instead of an element).”ynumber“Viewport y coordinate.”
- console_messagesreads or searchesdescribed as changing data
“Console messages and uncaught errors captured since this server attached to the tab.”
Publisher declares: read-only yes · open-world yes
5 inputs
clearboolean“Clear the buffer after reading.”levelstring“Only this level (error includes uncaught exceptions).”limitinteger“Last N entries (default 100).”patternstring“Regex the text must match.”tabstring“Tab id from tabs_list.”
“Run JavaScript in the page and return the JSON result.”
Publisher declares: read-only no · open-world yes
2 inputs, 1 required
codestringrequired“JavaScript expression or function body.”tabstring“Tab id from tabs_list.”
“Set values of several fields at once (inputs, textareas, selects by value or label, checkboxes/radios with true/false).”
Publisher declares: read-only no · open-world yes
2 inputs, 1 required
fieldsarrayrequired“Fields to set.”tabstring“Tab id from tabs_list.”
“Find visible elements whose text, label, placeholder, id or name contains the query.”
Publisher declares: read-only yes · open-world yes
3 inputs, 1 required
limitinteger“Max results (default 20).”querystringrequired“Text to look for.”tabstring“Tab id from tabs_list.”
“Return the visible text of the page (or of a CSS selector) — good for reading articles, results and messages.”
Publisher declares: read-only yes · open-world yes
3 inputs
max_charsinteger“Max characters (default 30000).”selectorstring“CSS selector to read (default body).”tabstring“Tab id from tabs_list.”
“Move the mouse over an element or coordinates (reveals menus/tooltips).”
Publisher declares: read-only no · open-world yes
6 inputs
matchstring“Visible text / accessible name to match (case-insensitive substring).”refstring“Element ref from snapshot/find, e.g.”selectorstring“CSS selector (first match).”tabstring“Tab id from tabs_list.”xnumber“Viewport x.”ynumber“Viewport y.”
“Navigate the tab to a URL, or go back/forward/reload.”
Publisher declares: read-only no · open-world yes
4 inputs
actionstring“History action instead of url.”tabstring“Tab id from tabs_list.”timeout_msinteger“Load timeout (default 30000).”urlstring“Destination URL; scheme defaults to https://”
- network_requestsreads or searchesdescribed as changing data
“Network requests captured since attaching to the tab, with status.”
Publisher declares: read-only yes · open-world yes
5 inputs
clearboolean“Clear the buffer after reading.”failed_onlyboolean“Only failed requests and HTTP status >= 400.”limitinteger“Last N entries (default 100).”patternstring“Regex the URL must match.”tabstring“Tab id from tabs_list.”
Show 11 more tools
“Press a key or combination on the focused element, e.g.”
Publisher declares: read-only no · open-world yes
3 inputs, 1 required
keysstringrequired“Key or combo; modifiers: Control, Shift, Alt, Meta.”repeatinteger“Times to press (default 1).”tabstring“Tab id from tabs_list.”
“Emulate a viewport size (e.g.”
Publisher declares: read-only no · open-world yes
3 inputs, 2 required
heightintegerrequired“Height in CSS px.”tabstring“Tab id from tabs_list.”widthintegerrequired“Width in CSS px.”
“Capture the visible viewport as PNG.”
Publisher declares: read-only yes · open-world yes
2 inputs
pathstring“File path to save the PNG to.”tabstring“Tab id from tabs_list.”
“Scroll the page (or a scrollable element) by direction/amount, or scroll an element into view.”
Publisher declares: read-only no · open-world yes
7 inputs
amountinteger“Pixels (default 600).”directionstring“Direction (default down).”matchstring“Visible text / accessible name to match (case-insensitive substring).”refstring“Element ref from snapshot/find, e.g.”selectorstring“CSS selector (first match).”tabstring“Tab id from tabs_list.”withinstring“CSS selector of a scroll container instead of the window.”
“Return a compact accessibility-style outline of the page: headings, landmarks and every visible interactive element with a ref (e.g.”
Publisher declares: read-only yes · open-world yes
4 inputs
max_charsinteger“Output budget (default 20000).”modestring“interactive (default) or all (adds text blocks).”selectorstring“Limit the snapshot to the subtree matching this CSS selector.”tabstring“Tab id from tabs_list.”
“Close a tab (default: current).”
Publisher declares: read-only no · open-world yes
1 input
tabstring“Tab id from tabs_list.”
- tab_newcontrols a browser or desktopdescribed as changing data
“Open a new tab (optionally at a URL) and make it current.”
Publisher declares: read-only no · open-world yes
1 input
urlstring“URL to open; scheme defaults to https://”
- tab_selectcontrols a browser or desktopdescribed as changing data
“Make a tab current (and bring it to front).”
Publisher declares: read-only no · open-world yes
1 input, 1 required
tabstringrequired“Tab id from tabs_list.”
“List open tabs.”
Publisher declares: read-only yes · open-world yes
Takes no inputs.
“Type text with real keyboard input.”
Publisher declares: read-only no · open-world yes
7 inputs, 1 required
clearboolean“Clear the field first (default true when an element is given).”matchstring“Visible text / accessible name to match (case-insensitive substring).”refstring“Element ref from snapshot/find, e.g.”selectorstring“CSS selector (first match).”submitboolean“Press Enter after typing.”tabstring“Tab id from tabs_list.”textstringrequired“Text to type.”
“Wait until text appears, a CSS selector is visible, or the URL contains a string (gone=true waits for the opposite).”
Publisher declares: read-only yes · open-world yes
7 inputs
goneboolean“Wait for the condition to become false instead.”msinteger“Plain sleep in milliseconds (max 60000).”selectorstring“CSS selector that should be visible.”tabstring“Tab id from tabs_list.”textstring“Text that should appear in the page.”timeout_msinteger“Timeout (default 10000).”urlstring“Substring the URL should contain.”
Each quoted sentence is the first sentence of the publisher’s own description. The kind of tool and the “changes data” mark are our AI model’s reading of that description, not something we tested: two differently worded runs had to agree, otherwise no label is shown, and nobody has checked them by hand. A tool without the “changes data” mark is described as only reading, searching or calculating, or the two runs did not agree. No tool was called. Read the publisher’s full description before relying on it. “Publisher declares” lists the hints the server itself sends with the tool (MCP annotations: read-only, destructive, idempotent, open-world); they are the publisher's claims, not checked by us. Inputs are the parameters in the tool's own input schema: name, declared type, whether it is required, and the first sentence of its description. Read 2026-10-23.
How well are its tools described?
Features #4 and #5
How clearly this server's own tool descriptions say what a tool does, and how consistent its tool list is. The first part is an AI model's yes/no reading of each description; the second part is counted. We do not turn it into a grade.
Descriptions
AI reading of each description
- Says what the tool does
- 100% of 23 tools
- Says what it returns
- 48% of 23 tools
- States limits, defaults or errors
- 43% of 23 tools
- Says when to use it
- 0% of 23 tools
- Items covered per tool
- 1.9 of 4 on average
Counted from the listing
- Parameters with a description
- 100% of 74
- Tools with a description
- 100% of 23
- Typical description length
- 112 characters
The tool set
Counted from the listing
- Tools
- 23
- Size of the list
- about 3,352 tokens for an agent to load (characters ÷ 4)
- Naming
- 13 of 23 names use snake_case
AI reading: easy to mix up
No pair found.
The AI items cover 23 of 23 tools: for a tool where two differently worded runs disagreed, that item is left out. They were not checked by hand, and no tool was called. “Easy to mix up” means two tools whose names or descriptions overlap without saying how they differ; the AI found none. We do not rate whether the tool set is complete or whether side effects are spelled out, because our AI reading of those was not reliable enough.
How do I add it to my client?
Config from the registry entry
- No config: the MCP bundle is a file to install in a client that supports bundles, not a command.
Built from the server's registry entry with each client's documented format (checked 2026-10-14; the Cursor link and the VS Code command 2026-10-19; the MCP Inspector command 2026-10-21). Values in < > are yours to fill in; this page never contains a key. Where the server needs a setting the entry does not declare, it can still fail to start.
Does the registry listing match what we found?
Feature #6The registry verifies a package once, when it is published, and does not re-check it. A difference below is about the listing, not a judgement of the publisher.
| Registry says | We found | Result | Source · date |
|---|---|---|---|
| Published package can be started | Started and answered | matches | our test · 2026-10-23 |
| Repository github.com/Gwakko/agent-in-browser | Opens publicly | matches | api.github.com · 2026-10-23 |
Who made it?
Feature #7 · lineage
Facts about the publisher. We do not turn them into a trust score.
- Namespace
- io.github.Gwakko
- GitHub owner
- a personal account (Gwakko)
- Repository
- github.com/Gwakko/agent-in-browser
- Stars
- 0
- Repository age
- 0–30 days
- Last push
- 0–30 days ago
- Archived
- no
- Security policy
- none found (we looked for SECURITY.md in the repository and in its owner’s .github repository)
Source: the registry entry and api.github.com · 2026-10-23; security policy read from GitHub · 2026-10-23.
Publish this server? Run these checks against it yourself before release, from your own machine: a server the command starts, or your own server at its URL after a one-time proof that you control it: npx protocolprobe-selftest.
Choosing
Record checks
Public records · not a safety rating4 of 5 record checks met · 4 do not apply
- Not ranked by downloads: not an npm or PyPI package
- Not ranked by stars: 0 stars (ranked from 10)
- MetRegistry entry is active and says how to run itActive, with a package or a remote address.
- MetSource repository link opensThe linked GitHub repository is public.
- MetRepository is not archivedNot archived.
- MetCode pushed to the repository in the 90 days before our checkLast push 0-30 days before our check.
- Not metRepository publishes a security policy (SECURITY.md)No SECURITY.md in the places GitHub reads it from.
- Does not applyPackage names its licence with a standard (SPDX) identifierNot an npm or PyPI package.
- Does not applyThe version the registry lists can be installed and is not withdrawnNot an npm or PyPI package.
- Does not applynpm package carries build provenance (attestation)PyPI packages have no npm-style attestation.
- Does not applyNo known critical or high advisory in the package or its dependenciesNot an npm or PyPI package.
Read from public records only: the registry entry, npm/PyPI, GitHub and advisory databases, checked 2026-10-23. No tool was called. Only “met” and “not met” are counted, so the total differs between servers; a check that does not apply or that we could not run is listed, never counted against it. It is not a safety rating. How each check works · Leaderboards
Common questions about io.github.Gwakko/agent-in-browser
Answered from the data aboveIs io.github.Gwakko/agent-in-browser safe to install?
We do not rate servers as safe or unsafe. 4 of its 5 applicable record checks are met (facts from public records, each listed on this page). What the records show: Its source repository opens publicly. Its repository had a push in the last 90 days. 16 of its tools describe changing data (our AI model's reading of the descriptions). These are facts with their sources on this page, not a safety guarantee, and we did not run its tools.
Has io.github.Gwakko/agent-in-browser been security tested?
Partly. 2 of the 12 checks that apply to it were exercised. We started the package in an isolated sandbox and read what it declares, but this test round was read-only, so its tools were not called. No security verdict is published before verification.
What can io.github.Gwakko/agent-in-browser do?
It lists 23 tools. In our AI model's reading of their descriptions (two runs had to agree, nobody checked it by hand), they include tools that read or search, create or change data, run code or commands and control a browser or desktop. We did not call any tool, so this describes the text, not the behaviour.
Can io.github.Gwakko/agent-in-browser change or delete data?
16 of its tools say in their description that they create, edit, delete, send or run something (our AI model's reading). We did not call any tool, so we cannot say what actually happens.
Does io.github.Gwakko/agent-in-browser publish a security policy?
We found no SECURITY.md in its repository (root, .github/ or docs/) or in its owner's .github repository.
Where is the source code of io.github.Gwakko/agent-in-browser?
In github.com/Gwakko/agent-in-browser, which opens publicly.
Found this page useful?
Anonymous · no sign-inA quick, anonymous reaction — no account, no email. One per kind from your browser. The tally updates when the page is next refreshed, and it is only a rough count, never a safety rating or a measured number of people.
Similar servers
Same topic · AI-read
- io.github.Agent360dk/browser-mcpAgent360 Browser MCPalso has tools that controls a browser or desktop, creates or changes data, runs code or commands, reads or searchescompare with this one
- io.github.imshaikot/browsenticBrowsenticalso has tools that controls a browser or desktop, creates or changes data, runs code or commands, reads or searchescompare with this one
- io.github.Silbercue/public-browserPublic Browseralso has tools that controls a browser or desktop, creates or changes data, runs code or commands, reads or searchescompare with this one
- io.github.protostatis/unbrowserunbrowser by Unchainedalso has tools that controls a browser or desktop, creates or changes data, runs code or commands, reads or searchescompare with this one
- io.github.Mingye-Lu/acrawlacrawlalso has tools that controls a browser or desktop, creates or changes data, runs code or commands, reads or searchescompare with this one
- io.github.mindstone/mcp-server-browser-automationBrowser Automationalso has tools that controls a browser or desktop, creates or changes data, runs code or commands, reads or searchescompare with this one
Active servers whose topic is also “search-and-web-data” and that have at least one tool of the same kind; most kinds in common first, then most GitHub stars. The topic and the tool kinds are an AI model's reading of the publishers' descriptions, not checked by hand. Similar is not a recommendation.
Compare with another server
Side by side: registry entry, our read-only session, the tools and the 18 classification fields. Same values as each page; no score.
Security testing
What did we test, and what did we leave out?
18 checks · features #1, #2
Not probed (read-only run)
We started the package in an isolated sandbox and read what it declares, but this test round was read-only, so its tools were not called. No security score is given; a check we did not run is never counted as passed.
- Inventory: exercised
- Hidden instructions: exercised
- Real execution: not run · our policy
- Secret leak: not run · nothing to check it against
- Internal addresses: not run · our policy
- Unconfirmed changes: not run · our policy
- Bad input: not run · our policy
- Model requests: not run · nothing to check it against
- Sensitive questions: not run · nothing to check it against
- Folder boundary: not run · our policy
- Log leaks: not run · nothing to check it against
- Hidden content: not run · nothing to check it against
- Fake assistant turns: not applicable
- Token passthrough: not applicable
- Token audience: not applicable
- Sign-in redirect: not applicable
- Consent handling: not applicable
- Sign-in metadata: not applicable
2 exercised10 not run (reason given)6 not applicable
Each check ends in exactly one state: exercised, not run (with the reason), or not applicable (with the reason). 2 of the 12 checks that apply to it were exercised. We do not publish security-check outcomes for named servers.
- 1Inventory
tool-listReads the declared tools, resources and prompts.exercised - 2Hidden instructions
description-poisoningLooks for hidden instructions in tool descriptions.exercised
- 3Real execution
real-executionCalls a tool and checks what really happens.This check calls the package's tools. This test round was read-only, so no tool was called; tools can act on outside services with the publisher's or a user's account.not run · our policy - 4Secret leak
canary-leakPlants a fake secret and checks whether it leaks.This check needs at least one successful tool call to have something to check, and no tool was called.not run · nothing to check it against - 5Internal addresses
ssrfChecks whether a tool can be steered to internal addresses.This check calls the package's tools. This test round was read-only, so no tool was called; tools can act on outside services with the publisher's or a user's account.not run · our policy - 6Unconfirmed changes
write-without-confirmationChecks whether tools that change data act without asking.This check calls the package's tools. This test round was read-only, so no tool was called; tools can act on outside services with the publisher's or a user's account.not run · our policy - 7Bad input
error-handlingSends bad input and watches how the server answers.This check calls the package's tools. This test round was read-only, so no tool was called; tools can act on outside services with the publisher's or a user's account.not run · our policy - 8Model requests
samplingChecks whether the server asks the client's model to act for it.This check needs at least one successful tool call to have something to check, and no tool was called.not run · nothing to check it against - 9Sensitive questions
elicitation-sensitive-dataChecks whether the server asks users for sensitive data.This check needs at least one successful tool call to have something to check, and no tool was called.not run · nothing to check it against - 10Folder boundary
root-boundaryChecks whether the server reaches outside the folders it was given.This check calls the package's tools. This test round was read-only, so no tool was called; tools can act on outside services with the publisher's or a user's account.not run · our policy - 11Log leaks
logging-side-channelChecks whether log messages carry data they should not.This check needs at least one successful tool call to have something to check, and no tool was called.not run · nothing to check it against - 12Hidden content
audience-hidingChecks resources that hide content from the user.This check needs content the server returns, and nothing was read or called.not run · nothing to check it against
- 13Fake assistant turns
fake-assistant-turnChecks prompts that pose as the assistant's own words.The server exposes no prompts.not applicable - 14Token passthrough
token-passthroughChecks whether the server forwards a user's token to other services.The server runs locally over standard input/output, so there is no sign-in flow to check.not applicable - 15Token audience
token-audience-validationChecks whether tokens meant for another service are accepted.The server runs locally over standard input/output, so there is no sign-in flow to check.not applicable - 16Sign-in redirect
open-redirectChecks the sign-in redirect for abuse.The server runs locally over standard input/output, so there is no sign-in flow to check.not applicable - 17Consent handling
confused-deputyChecks consent handling when one service acts for another.The server runs locally over standard input/output, so there is no sign-in flow to check.not applicable - 18Sign-in metadata
oauth-url-schemeChecks the published sign-in metadata for unsafe URLs.The server runs locally over standard input/output, so there is no sign-in flow to check.not applicable
Source: our own test in an isolated sandbox, 2026-10-23, mcp-runtime-engine@0.1.0. Which checks count and why follows our published scoring model.
What happened during the test?
Feature #8
The steps in order, as the test record holds them.
- 0 msStarted the published package in an isolated sandbox (read-only mode)
- 30 msConnected after 0 redirects.
- Read the inventory: 23 tools, 0 resources, 0 prompts.
- 1,760 msFinished: 2 checks exercised.
Reference
How is it classified?
18 fields
One value per field, each with its source. Fields written by AI models are marked. A field with no value says why.
“Not stated” means the source did not state it; “does not apply” means it does not apply to this kind of server; “our limit” means our own lookup failed. A linked value opens the directory filtered to the same value.
Cite or correct this page
Sources and corrections
Suggested citation
ProtocolProbe. “io.github.Gwakko/agent-in-browser: MCP server record.” Data checked 2026-10-23. https://protocolprobe.com/mcp/servers/io.github.Gwakko/agent-in-browser
Figures are valid for the date shown. Our facts on this page are CC BY 4.0: reuse them with a link and the check date. Quoted text (the publisher’s descriptions, website quotes, tool names) is not ours to license.
As JSON: this record in the API · API documentation.
Badge for your README
[](https://protocolprobe.com/mcp/servers/io.github.Gwakko/agent-in-browser)
It shows this server's record-check count: facts from public records, no tool called. It is not a security rating or a pass. It links here and updates after each data reload.
Something wrong, or is this yours?
If a fact here is wrong, tell us and we will re-check it. If you operate this server, you can run our checks on it yourself with the free self-test command; we call no server’s tools ourselves for anything shown here.
Write to hello@protocolprobe.com · corrections log · methodology