Repository navigation
Releases
WebAIAssist 0.1.0
First release. Sent to the Chrome Web Store for review on the same day.
Added
-
Stop ends a model that does not stop: when a graphics-card model is still running five seconds after
Stop, its engine is ended and the panel says to press Start again (a model that runs in a Worker is terminated;
one on the page's main thread cannot be, and the next Start waits for it). A model too
slow for the graphics card could otherwise keep the browser's graphics process busy for minutes. -
The agreed remote hosts are read when the question is due, not when the panel opened: after "Forget
all" or "Reset to defaults" an open panel asks again before it uses a remote chat service, and a remote
model that was already started is let go before the next question. -
Reset to defaults keeps the host access of a site you allowed, when the chat endpoint is on the same host.
-
A node page without tools says what to do next: open an item page of the site for the node's tools, or
press Read this page to ask about the open page (the old line put both in a bracket). -
A forgotten site in an open tab: the page scripts left in a tab whose site was forgotten are no longer
asked for its tools; the panel shows Allow again. -
The Activity page has one heading and no longer shows the graph page's sentence about a node's files.
-
Default models from the measured runs (15 and 16 Nov 2026): Chrome's built-in model when Chrome has it
ready; otherwise Qwen3 0.6B on the graphics card whenever the browser can use it (a slow first answer is named in
the reason line); otherwise Qwen3 1.7B on the processor. Llama 3.2 3B and Gemma 4 E2B stay in the graphics-card list, each marked "slow: may stall on
a laptop graphics card" and never picked by default: on the measured graphics card both ran past five minutes on
their first question. -
A page that registers its tools late: when the first list of a page is empty, the panel lists it again
after one and after three seconds; the toolbar number is read again as soon as you allow or forget a site.
Before, both could stay at "no tools" until the page changed something (seen on the first case of the
conformance page in Chrome 157). -
Build checks in the release gate:
npm run test:distruns the checks of the built package and the
privacy host scan after the build and the zip, and fails when the build is missing, so CI no longer skips them. -
Release workflow in two jobs: the build runs with read access only; the job that creates the release runs
no installed code and attests the zip's build provenance (gh attestation verify). -
Locale file: the build writes
_locales/en/messages.jsonfrom the typed English dictionary, and the
description and the toolbar button title in the manifest come from it (default_locale: en), so another
language is one more dictionary and one more folder. -
Tool security review and parity features: nine written rules read each tool's definition (hidden
characters, orders to the AI, a call to another tool, look-alike names, links, encoded text, a read-only mark
that the name contradicts, oversized text, a definition changed since first seen); a tool with a finding is
marked "check", shows its reasons in the list, the detail view and the question, and is always asked about
(docs/TOOL-REVIEW.md). A tool detail view with every hint, its fields and its reasons;
"Run it by hand" through the same questions, written to the activity log as an entry with no model; a second question for a tool the page marks as consequential;
form tools say that the page waits for you to press Submit and warn about password, payment and personal
fields; tools with thedebugginghint are listed as "for developers" and are not offered to the model; a number
on the toolbar icon counts the tools of a site you allowed; "Copy report" lists the review's findings. Stop now
reaches the tool that runs on the page, a result of nothing reads as one plain sentence, a same-origin frame's
tools are neither listed nor run, and a call whose page navigated is not sent again. The built-in graph tool
asks before each call in "ask for every call". -
Node card, help screens and Copy report: node card heading from the node's own link set
and a "Start your own node" command to copy; a fourth starter ("Compose a harness for my task")
and anexplainstarter that asks why two items are linked; help screens for "This site has no
tools" and "Chrome needs the WebMCP switch"; a developer "Copy report" of the open page's tools
with a "Security review flags" hook (reviewFlags, empty for now). -
Side panel: a chat with a local AI model that uses the WebMCP tools of the open page.
Streaming answers, Stop, a step limit (1 to 12, default 6), Export as Markdown (it says it was
written by an AI model and names the model and where it ran), Help and a welcome screen,
keyboard use (Ctrl+Enter, Escape), light and dark themes, no library, font or image from
outside the package. -
Four model paths behind one
ModelRuntimeport:- the graphics card: WebLLM (
@mlc-ai/web-llm0.2.86), Qwen3 0.6B, Qwen3.5 2B, Llama 3.2 3B and Gemma 4 E2B, in a Worker
of the offscreen document (or in the service worker when the offscreen document has no
WebGPU); - the processor: llama.cpp in WebAssembly (
@wllama/wllama3.8.1), Qwen3 0.6B, Qwen3 1.7B and
Qwen3.5 2B as GGUF
files with a recorded size and SHA-256, checked after the download, stored once, Chrome 137 or
newer; - Chrome's built-in model (Prompt API), with its context window read from Chrome, a compact tool
list, replies constrained to the action schema, and a plain reason for each state; - an OpenAI-compatible chat endpoint (Ollama and LM Studio presets;
https, orhttpon
localhostand127.0.0.1only; no redirects, replies capped at 1 MB).
- the graphics card: WebLLM (
-
Chat endpoint, local or remote: an OpenRouter preset and "Other address"; the model list
offers "Use your own server or a remote service…" while no endpoint is set; a one-time question
per host before a remote service is first used (where the model runs, what is sent, that it may
charge; nothing is sent before Use this service; never forlocalhost); the endpoint
runtime reads token counts and the price from the streamed reply (onUsageon the chat
request); a service's error text is cut and scrubbed of the API key;npm run eval:endpoint
runs five questions against any endpoint (address, model and key from the environment only).One model is in memory at a time. The model libraries and the llama.cpp engine ship inside the
package, pinned by SHA-256; only weights download, from fixed Hugging Face addresses. -
Default model by measurement: Chrome's built-in model when ready; the graphics card when its
first answer was not slow (20 seconds); the processor model; then an endpoint. The panel says
why in one sentence. A model the user chooses is kept. Options has Automatic. -
Consent before a download: size, source and place of storage are shown, and nothing downloads
before Download. Options lists the downloaded models with sizes and removes one or all.
Forget all sites clears everything kept per site. -
Where each model runs is stated everywhere: the model list has two groups ("On this computer,
in the browser" and "Your own server or a remote service"), every label and the model line after
Start say whether the model runs in this browser on the graphics card, in this browser on the
processor and memory, in Chrome on this computer, on your own server on this computer, or on a
remote service (then also where the chat goes), and Options says the same. -
Per-model prefill chunk for the graphics-card models (a bigger model prefills in shorter
chunks, so one chunk stays under the graphics driver's time limit), and plain handling of a
model that is gone: a lost device, or an engine WebLLM no longer has loaded, stops the model
line saying "Ready" and shows plain words. -
Page card that follows the browser: it refreshes by itself when the tab, window, address,
title or the page's tools change, and shows the site, an access badge (No page, Cannot run here,
Not allowed yet, Allowed, Blocked by the site), each WebMCP tool with the page's own words and
"reads" or "changes", and a More menu (Add WebMCP, Forget, Export). A tool with no hint counts as
"changes". -
Site information: for a site the user allowed, a fixed list of public files is read from the
same site (llms.txt,llms-full.txt, the knowledge link set, API catalog, agent card,
security.txt, the AI policy ofrobots.txt, TDMRep) and the page's JSON-LD. Same origin, GET,
no cookies, no redirects, 5 and 10 second limits, size caps, at most eight files. Shown as
published by the site; a few short facts go to the model as untrusted data (Options can turn
this off). A site that sendsContent-Signal: ai-input=nogets no site text sent to the model
and no Read this page, unless the user chooses Use on this tab anyway (local model only). -
AgenticSystemCore node mode: a site that says it is a node is shown as saying so, and the
model gets a short hint on the roles of the ten node tools. -
Scope of action: only the open page's own WebMCP tools, through the confirmation gate (tool,
site, arguments; Allow once or Deny; once per site per session only for read-only tools, in
Options). The tool is checked again before the call and after the question. -
Agent loop: one action per reply, one repair try that names the tool's fields, a call
identical to an earlier one is not run again, untrusted-data fence for tool results, a run bound
to the tab and site it started on. -
Options page: where the model runs, model per path, endpoint, confirmation mode, step limit,
memory (and Clear memory), si...