Jev × WebMCP
jev-webmcp-extension (github.
Overview
jev-webmcp-extension (github.com/sdras/jev-webmcp-extension) is Sarah Drasner's Apache-2.0 Chrome side panel. It discovers tools a page registers, converts each schema field into parallel Jev questions (tool Choice plus per-argument Choice, Noul, or span picks from the user's words), and decodes answers into executable calls with confidence and latency labels. Manifest screening runs Noul checks on tool descriptions for agent-directed instructions; execution policy respects readOnlyHint, consequential annotations, and double-Enter confirmations. No build step: load unpacked on Chrome 149+ with WebMCP enabled. Chrome Web Store listing and Basketful demo walkthrough ship in README; eval harness hits the real TypeSafe API when TYPESAFE_API_KEY is set.
Problem: WebMCP pages expose many typed tools, but natural-language commands need reliable tool and argument selection without site-specific prompt hacks.
Built for: Frontend and agent builders experimenting with Chrome WebMCP who want Jev Choice and Noul batches mapped from JSON Schema at typing speed.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- Sarah Drasner
- Handle
- @sdras
How Jev is used
- Role in the product flow
- Side-panel tool routing and argument filling from schema-derived Jev question batches
- Primitives
- ChoiceNoul
- State in
- User utterance plus discovered tool manifests converted to questions in src/core/questions.js (tool list Choice, enum booleans, optional-field nouls, span Choices for free text).
- Decision out
- Selected tool name, argument object, confidence score, and policy state (auto, confirm, or none).
- Discover tools via page bridge and screen manifest text
- Build parallel Jev questions and decode plan for the utterance
- Call api.typesafe.ai/v1/systemone through extension storage key
- Apply policy.js gates before chrome.scripting executes pageCallTool
Drasner's extension is the browser-local counterpart to server-side Browser Use loops: instead of pixels, Jev chooses among declared tools and fills args using the user's own words as candidate spans. That keeps latency in the hundreds of milliseconds for read-only calls while still forcing confirmations on checkout flows. Core logic lives in pure JavaScript modules tested without Chrome so question wording tweaks are safe to iterate. Tool results stay out of model input per security README, which matters when pages are untrusted. Pair with typesafe-ai-typesafe-sdk-js when you graduate from the panel to a hosted agent backend.
Sourced performance claims
- README example shows search_products at 98% confidence in 164 ms on Basketful demo input.Source: github.com/sdras/jev-webmcp-extension README
- Public GitHub repo sdras/jev-webmcp-extension had about one hundred nine stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- Schema-to-question converter with npm test coverage
- Per-site optional_host_permissions and manifest screening badges
- Keyboard navigation through tool candidates with Enter to execute
- Open in playground link for debugging System One requests
Stack
- JavaScript
- Chrome extension APIs
- WebMCP
- TypeSafe System One HTTP
Pricing: Free extension; TypeSafe API keys bill per System One request when not using harness mocks.
Links
FAQ
- Do I need WebMCP enabled?
- README requires Chrome 149+ with WebMCP via origin trial or chrome://flags/#enable-webmcp-testing before tools appear.
- Can any site auto-run destructive tools?
- Policy treats tools as state-changing unless readOnlyHint is set; consequential annotations always need a second Enter per README safety section.
- Where is the TypeSafe key stored?
- Panel settings save the key in chrome.storage.local after a live API check per setup steps.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.