Listing on LM Studio Hub:
webbrain/web-tools(ownerwebbrain, nameweb-tools, public) — https://lmstudio.ai/webbrain/web-toolsOne-click install while LM Studio is open: click the Run in LM Studio button on the Hub page, or open this URL on your machine:
lmstudio://plugin?owner=webbrain&name=web-toolsThat triggers LM Studio to download the plugin from the Hub into your local plugins directory. You can also install from inside LM Studio (integrations icon next to the chat input → toggle webbrain/web-tools on). For a CLI install, clone the artifact, enter the cloned directory, then install it:
lms clone webbrain/web-tools cd web-tools lms dev --install
Give any LM Studio model the ability to read the live web — and, if you have the browser extension, to act inside your own signed-in session.
Works without the browser extension:
fetch_url— raw HTTP fetch with content-type smarts. JSON gets pretty-printed, HTML is stripped to readable text +<title>, plain text comes back verbatim unless it is long enough to compact, binaries are summarised instead of inlined.research_url— same fetcher, biased toward "give me the article body, not the navigation chrome." Extracts<main>/<article>, drops header/nav/footer/aside before stripping tags. Best for news, blog posts, READMEs, Wikipedia, docs.
Adds real browser access when the WebBrain extension is installed on a Chromium browser (Chrome, Edge, Brave, Opera, Vivaldi — not Firefox):
browser_task— hand a goal to WebBrain running in the browser you are already logged into. Reaches the authenticated dashboards, webmail and client-rendered apps that plain HTTP cannot see.mode='ask'is read-only;mode='act'can click and type, gated by in-browser approval prompts.browser_status— report whether the extension is attached, and what to do if it isn't. Pass arunIdto poll a task that outlived its initial timeout and retrieve the eventual result.browser_respond— forward the user's answer when a task pauses withneeds_user_input. Use therunIdandclarifyIdreturned by the task; never guess the answer on the user's behalf.browser_abort— stop a continuing run byrunId. Browser actions that already completed are not rolled back.
The browser tools are always offered. Without an extension they return an actionable message instead of failing opaquely, so the model can tell you how to switch the capability on rather than retrying blindly.
Long text results are context-friendly by default. Instead of dumping
the first N characters and losing everything after the cutoff, the tools
run an LLM-free compaction pass that keeps the beginning, a few high-signal
middle passages, and the ending. Results include compacted:true plus a
compaction object when this happens. Pass compact:false for legacy
head-only truncation, or maxChars to ask for a smaller/larger return
budget (fetch_url: 8k text / 16k JSON by default, hard-capped at 50k;
research_url: 16k by default, hard-capped at 60k).
Pure Node — no Puppeteer, no Playwright, no headless Chromium.
Same fetching logic the WebBrain browser
extension ships, ported off chrome.* onto Node fetch.
The simplest installation path is the Run in LM Studio button on the
Hub listing. For a terminal-based
install, lms clone only copies the artifact into the current directory; run
the install command from that cloned plugin directory:
lms clone webbrain/web-tools
cd web-tools
lms dev --installOpen LM Studio, click the integrations icon next to the chat input, and toggle
webbrain/web-tools on. Any tool-capable model in the same chat can now call
fetch_url and research_url.
Try it:
What's on the front page of news.ycombinator.com right now?
The model should call research_url and come back with live
headlines.
browser_task needs the WebBrain extension on a
Chromium browser — Chrome, Edge, Brave, Opera or Vivaldi. The extension
dials out to this plugin; a Manifest V3 extension cannot listen on a socket,
so the plugin hosts the listener.
The bridge runs from the extension's offscreen document, and the Firefox
build has none — see src/firefox/ARCHITECTURE.md.
cloud-bridge.js and cloud-runs.js exist only under src/chrome/. On
Firefox, browser_task will always return the not-connected response.
fetch_url and research_url are unaffected and work everywhere.
- Install the extension and open your browser.
- Enable the plugin once so it creates its pairing key, then copy the key
from
~/.webbrain/bridge-key(%USERPROFILE%\.webbrain\bridge-keyon Windows). The file is readable only by you. If you also use the MCP server, it is the same key (npx -y @webbrain/mcp-server pairing-keyprints it). - In WebBrain → Settings → Bridge, set the URL to
ws://127.0.0.1:17375/extension, paste the key into Pairing key, and enable it. - Ask the model to call
browser_statusto confirm.
One bridge at a time. The extension holds exactly one outbound bridge socket. Pointing it here means it is not pointed at WebBrain Cloud (
17373) or the MCP server (17374). Override the port withWEBBRAIN_BRIDGE_PORT, and the key withWEBBRAIN_BRIDGE_KEYorWEBBRAIN_BRIDGE_KEY_FILE.
Upgrading from an unpaired version. Earlier plugin versions accepted any
socket that sent the right hello. Update the WebBrain extension first (a
version without the Pairing key field cannot pair and is disconnected after
10 seconds), then paste the key as in step 3. "Rejected by the backend: Pairing
key mismatch" under Settings means the pasted key is not the plugin's.
Try it:
Check my GitHub notifications and summarise anything that mentions me.
Listed up front so it doesn't bite you mid-session. All three limits apply
to fetch_url and research_url only — browser_task is the answer to
every one of them:
- No JavaScript rendering. Single-page apps that hydrate from
JSON (Twitter, Notion, modern dashboards) return near-empty text.
The plugin flags this with
spaSuspected: trueso the model can give up gracefully instead of looping. - No clicks, no typing, no screenshots. Those need a real browser.
- No cookies, no login. Each fetch is anonymous from your IP with no shared session — pages behind a login are out of reach.
If you want to hack on the plugin, clone this repo (rather than
lms clone the published artifact) and:
npm install
lms devlms dev bundles src/ and registers the plugin live with your
running LM Studio instance — edits to TypeScript files hot-reload
without a restart. The dist/ build via npm run build is only
needed if you want to import the tool implementations from another
Node project.
.
├── manifest.json ← LM Studio plugin metadata
├── package.json ← npm deps + build scripts
├── tsconfig.json ← strict-mode TS, ES2022
└── src/
├── index.ts ← plugin registration glue
├── tools/
│ ├── fetchUrl.ts ← fetch_url implementation
│ ├── researchUrl.ts ← research_url implementation
│ └── browserTask.ts ← browser task + status/respond/abort recovery tools
└── util/
├── htmlToText.ts ← regex HTML stripper
├── safeFetch.ts ← redirect-revalidating wrapper + streaming cap
├── urlGuard.ts ← private-IP / file:// blocker
├── bridgeClient.ts ← WebSocket listener the extension dials into
└── pairing.ts ← pairing key file and challenge proof
bridgeClient.ts and pairing.ts are deliberate standalone copies of the same
protocol in mcp-server/src/bridge.ts and
mcp-server/src/pairing.ts. The two ship to
different registries — LM Studio Hub and npm — and sharing a package would
drag the monorepo into both installs. Change the protocol in one, change
it in the other.
tools/ and util/ are pure functions with no SDK dependency — you
can import them from any Node project that wants the same
web-fetching primitives.
The plugin runs in your local Node process, so anything reachable from that process is reachable from the LLM. The defenses below are layered — each closes a class of bypass the previous one missed — but they are not a hardened sandbox (see end of this section).
The browser bridge is a separate trust surface from the HTTP tools.
Everything below concerns fetch_url / research_url. For browser_task:
-
The bridge listener binds
127.0.0.1only. Anything that can reach that port can drive your signed-in browser — never expose it to a network or a container bridge. -
Connections must send the extension's
helloframe and answer a pairing challenge: the plugin sends a random nonce and the extension returns an HMAC-SHA256 of it keyed with the pairing key, so the key never crosses the socket. Until a socket has done both it receives no commands and cannot replace the connected extension; a wrong answer closes it with code4001. The protocol is the MCP server's; see its Pairing section. -
browser_taskdelegates a goal, never individual clicks. WebBrain's capability × origin permission gate runs inside its own agent loop, so every approval prompt a human would see still fires. That is why this plugin does not expose the low-level browser primitives. -
A
timeoutdoes not cancel the run. A task that already submitted a form should not be silently killed; the browser keeps going and the result stays visible in the WebBrain side panel. -
URL guard, structural (sync). Requests to RFC1918 (
10.*,172.16-31.*,192.168.*), loopback (127.*,::1), link-local (169.254.*,fe80::/10), unique-local IPv6 (fc00::/7), carrier-grade NAT (100.64.0.0/10), unspecified (0.0.0.0/8,::), multicast, cloud-metadata IPs, IPv6 addresses that embed any of those IPv4 ranges (IPv4-mapped::ffff:7f00:1in any spelling, IPv4-compatible, NAT6464:ff9b::/96, 6to42002::/16, Teredo), and*.local/*.internal/*.lan/*.home/*.corp/*.intranetare blocked. Non-http(s)protocols (file://,ftp://,gopher://,javascript:) too. -
DNS resolution check, pinned to the connection. On each hop the hostname is resolved by the socket's own
lookuphook viadns.lookup({all: true}); every returned A/AAAA address is checked against the same private ranges, and only those checked addresses are dialled. Closes the bypass whereattacker.example A 127.0.0.1would otherwise pass the syntactic check, and DNS rebinding — a TTL=0 record cannot answer public to the check and private to the connection, because they are the same lookup. Requests use a fresh connection each time, never a pooled socket resolved earlier. -
Per-redirect re-validation. Redirects are followed manually (5 hops max). Each
Locationruns through both checks before we follow it — a public URL cannot 302 tohttp://169.254.169.254/.... -
Cross-origin header stripping. When a redirect crosses origins,
Authorization,Cookie,Proxy-Authorization, andProxy-Authenticateheaders are dropped. Stops credentials passed via the per-callheadersarg from leaking through an open-redirect chain. -
Streaming response cap. Bodies read with a hard byte ceiling (4 MB for
fetch_url, 6 MB forresearch_url). Past the cap, the stream is cancelled and the result is markedtruncated: true. -
No
credentials: 'include'. Unlike the browser extension, this plugin makes anonymous requests. -
Timeouts. Default 30 s, hard cap 120 s.
Private targets are opened by you, not by the model. To let fetch_url
and research_url reach localhost or your LAN, turn on Allow private
network access in this plugin's settings in LM Studio. The setting is
per chat and off by default. The tools have no allowPrivate argument,
and one sent anyway is ignored, so a page the model reads cannot talk
it into switching the guard off.
Cloud metadata endpoints stay blocked even with the setting on:
169.254.169.254, 169.254.170.2, 169.254.170.23, 169.254.0.23,
100.100.100.200, fd00:ec2::254, fd00:ec2::23,
metadata.google.internal, metadata.goog and metadata, including
IPv4-mapped IPv6 spellings, hostnames that resolve to them, and
redirects to them.
While the setting is on, the model still chooses the URL, method, headers and body, and a page it reads can steer it toward your router, admin panels or local services. Turn it on only in chats that need it.
These checks only see the address ranges they know about. If you're
running this on a cloud VM with sensitive instance-metadata
endpoints, use a network policy that blocks 169.254.169.254
outbound rather than relying solely on this plugin's checks.
This independently published plugin remains MIT-licensed. See LICENSE.
