{"name":"typesafeai.app capability feed","unofficial":true,"generatedFrom":"validated Git content","lastVerifiedAt":"2026-09-18","count":16,"sourceTypes":{"official":3,"community":13},"sourceTypeDefinition":{"official":"Published by TypeSafe itself.","community":"Published by a third-party person or organization."},"capabilities":[{"slug":"search-flights-in-seconds","title":"Search Google Flights in seconds","summary":"A browser agent turns the visible DOM into an indexed action space and uses Jev to choose the next operation and compatible target.","whatJevDoes":"Jev receives a compact table of visible controls and answers an operation Choice plus speculative target Choices in one request. Code reads only the target that matches the selected operation, resolves it back to an observed DOM node, checks freshness and occlusion, and executes it. A separate small text model is called only when the chosen operation needs free text.","integrationPattern":"Browser state to parallel operation and target choices; deterministic code validates and executes one branch.","jobs":["control","route"],"domains":["agents","automation"],"primitives":["choice"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/browser-use/jev-ultrafast","author":"browser-use","externalId":"browser-use/jev-ultrafast","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"metric":{"name":"Google Flights task time","value":7.073,"unit":"seconds","conditions":"Author-reported single recorded Zurich-to-London search; loading and text generation included.","sourceUrl":"https://github.com/browser-use/jev-ultrafast","independentlyVerified":false},"modelVersion":"jev-latest","limitations":["The published timing covers one browser profile and one task, not a general reliability benchmark.","Frames, canvas, uploads, pop-up tabs, and arbitrary keyboard widgets remain outside the MVP."],"repoUrl":"https://github.com/browser-use/jev-ultrafast","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"drive-macos-from-structured-screen-state","title":"Drive macOS from structured screen state","summary":"A computer-use loop combines OCR and accessibility data, then asks Jev which bounded action should move the Mac toward a plain-English goal.","whatJevDoes":"Jev classifies the next action and, when relevant, the visible or off-screen target from a numbered screen inventory. Capture, OCR, accessibility traversal, date arithmetic, action execution, and stop conditions remain deterministic. A writing model is isolated to free-text fields and proposed URLs, and password fields are refused.","integrationPattern":"OCR and accessibility observations become parallel Choices; code owns permissions, safety stops, and input execution.","jobs":["control","classify"],"domains":["agents","automation"],"primitives":["choice","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/awlevin/typesafe-computer-use","author":"awlevin","externalId":"awlevin/typesafe-computer-use","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"metric":{"name":"Decision cost","value":0.0002,"unit":"USD per step","conditions":"Author-reported comparison on the same screenshot and goal; excludes the optional writing model.","sourceUrl":"https://github.com/awlevin/typesafe-computer-use","independentlyVerified":false},"modelVersion":"jev-latest","limitations":["It targets macOS and requires Screen Recording and Accessibility permissions.","OCR and accessibility coverage miss canvas content, icon-only controls, and applications with weak accessibility trees."],"repoUrl":"https://github.com/awlevin/typesafe-computer-use","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"quote-a-market-every-block","title":"Choose a market quote every Monad block","summary":"A trading loop reads the Kuru MON-USDC order book and asks Jev for a buy-or-sell judgment before code places a post-only limit order.","whatJevDoes":"Jev evaluates the current order-book state and chooses buy or sell for a configured horizon. Ordinary code reads the book, applies position and margin caps, sets the price and size, signs the transaction, tracks receipts, and accounts for fills and P&L. A mock heuristic is the default, so a real Jev run requires explicit configuration.","integrationPattern":"Order-book snapshot to one bounded direction Choice; deterministic trading code constrains and submits the quote.","jobs":["control"],"domains":["trading","automation"],"primitives":["choice"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/jarrodwatts/jev-trader","author":"jarrodwatts","externalId":"jarrodwatts/jev-trader","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["The public deployment described by the repository is a dry run using a mock model, not evidence of live Jev profitability.","A directional judgment does not control transaction inclusion, fills, market movement, or financial risk."],"repoUrl":"https://github.com/jarrodwatts/jev-trader","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18"},{"slug":"choose-drone-tactics-from-camera-state","title":"Choose drone tactics from camera-derived state","summary":"A simulated quadrotor uses Jev for low-frequency tactical judgments while classical vision, flight control, and safety reflexes remain in code.","whatJevDoes":"Jev reads symbolic scene features derived from onboard depth and segmentation buffers. It chooses a maneuver, scores risk, and estimates whether a target is truly lost. A 50 Hz reflex layer can veto those judgments, while a 500 Hz geometric controller owns flight dynamics; Jev never receives pixels or direct motor control.","integrationPattern":"Classical perception to Choice, Score, and Noul judgments; faster deterministic loops retain safety authority.","jobs":["control","score"],"domains":["robotics"],"primitives":["choice","score","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/RomanSlack/jev-drone","author":"RomanSlack","externalId":"RomanSlack/jev-drone","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"metric":{"name":"Obstacle-course distance","value":77.5,"unit":"meters","conditions":"Author-reported single Jev run in MuJoCo; the matched multi-seed result in an earlier arena showed no advantage.","sourceUrl":"https://github.com/RomanSlack/jev-drone","independentlyVerified":false},"modelVersion":"jev-latest","limitations":["The strongest course result is one simulation run and does not establish a repeatable improvement.","Perception and safety are separate code paths; the model neither sees camera pixels nor controls motors directly."],"repoUrl":"https://github.com/RomanSlack/jev-drone","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"play-super-mario-from-emulator-state","title":"Play Super Mario from emulator state","summary":"An experimental controller translates NES telemetry into object-centric JSON and lets Jev choose the next legal controller macro.","whatJevDoes":"Each request asks Jev to select a controller action, estimate whether a forward jump is useful, and score immediate danger. The harness calculates timing facts, parses RAM, advances the emulator, and records outcomes. The repository provides a runnable harness but does not publish a completed-level claim or success rate.","integrationPattern":"Emulator telemetry to bounded controller Choice with companion jump and danger judgments.","jobs":["control","score"],"domains":["games"],"primitives":["choice","score","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/fhshaik/typesafe-mario","author":"fhshaik","externalId":"fhshaik/typesafe-mario","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["The repository demonstrates the controller architecture but does not report a verified level-completion result.","Users must provide a lawful local emulator and game setup; no Nintendo assets are included."],"repoUrl":"https://github.com/fhshaik/typesafe-mario","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18"},{"slug":"route-code-review-investigations","title":"Route focused code-review investigations","summary":"A staged review workflow uses Jev to identify risky areas, select evidence, classify mechanisms, score severity, and route follow-up checks.","whatJevDoes":"Jev supplies bounded risk, file, evidence, mechanism, severity, and reviewer-routing judgments across a sequence of focused calls. Code discovers Git changes or repository files, enforces workflow order, persists reports, and serves a local dashboard. The resulting findings are prompts for human review rather than proof that a defect exists.","integrationPattern":"Broad risk screening narrows into evidence selection and severity scoring under a code-owned review workflow.","jobs":["verify","route","score"],"domains":["developer-tools"],"primitives":["choice","score","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/devagrawal09/jev-review","author":"devagrawal09","externalId":"devagrawal09/jev-review","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["It does not integrate compiler diagnostics, static analyzers, or repository indexing.","Model-selected findings require independent review and do not prove that a defect exists."],"repoUrl":"https://github.com/devagrawal09/jev-review","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18"},{"slug":"complete-a-starcraft-shareware-mission","title":"Complete a StarCraft shareware mission","summary":"A reproducible harness lets Jev direct combat, exploration, and economy actions in the original StarCraft shareware campaign.","whatJevDoes":"Jev receives owned and currently visible game state, selects an intent category, and chooses a compatible command inside that category. The harness translates choices into ordinary mouse and keyboard input, verifies command acceptance, preserves action probabilities, and records evidence. The original game remains responsible for simulation, enemies, combat, and victory.","integrationPattern":"Observed game memory to hierarchical intent and command Choices; a deterministic adapter executes legal input.","jobs":["control","route"],"domains":["games"],"primitives":["choice"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/phyous/tsai-sc","author":"phyous","externalId":"phyous/tsai-sc","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"metric":{"name":"Verified development-run outcome","value":1,"unit":"mission completed","conditions":"Author-reported attempt 16 with evidence bundle and visually reviewed victory screen; not a win-rate benchmark.","sourceUrl":"https://github.com/phyous/tsai-sc","independentlyVerified":false},"modelVersion":"jev-1.13.0","limitations":["One successful development run does not establish a win rate or competitive real-time performance.","The game is paused for state reads and inference, and action probabilities are not probabilities of winning."],"repoUrl":"https://github.com/phyous/tsai-sc","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"supervise-a-coding-agent","title":"Supervise a coding agent while it works","summary":"Foreman runs an independent observation loop that asks Jev whether a coding worker is progressing, stuck, complete, or ready for verification.","whatJevDoes":"Jev evaluates nine narrow yes-or-no dimensions over bounded factory evidence, including requirement coverage, test sufficiency, progress, drift, and human need. A deterministic policy converts those probabilities into permitted interventions such as continue, verify, retry, finish, or escalate. Codex remains responsible for software-engineering reasoning and tool use.","integrationPattern":"Bounded worker evidence to parallel Noul assessments; a safety-ordered policy chooses the allowed intervention.","jobs":["verify","control","score"],"domains":["agents","developer-tools"],"primitives":["noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/thruwire/foreman","author":"thruwire","externalId":"thruwire/foreman","access":"public","publishedAt":"2026-09-17","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["Assessment accuracy and policy thresholds are not calibrated for production use.","The first version runs one coding worker at a time and local execution is not isolated."],"repoUrl":"https://github.com/thruwire/foreman","permissionStatus":"public-source","publishedAt":"2026-09-17","lastVerifiedAt":"2026-09-18"},{"slug":"gate-coding-agent-tool-calls","title":"Gate risky coding-agent tool calls","summary":"A Pi extension asks Jev to flag destructive, exfiltrating, or out-of-scope tool calls and to classify failures in command output.","whatJevDoes":"Before bash, write, or edit calls, Jev scores destructive impact, exfiltration, and scope risk. After selected command results, it checks for leaked credentials and classifies failures so code can append fixed advice. Shadow mode is the default, all error paths fail open, and enforcement requires explicit configuration.","integrationPattern":"Tool intent and bounded output to parallel risk judgments; extension code applies thresholds and user-confirmation policy.","jobs":["verify","classify","score"],"domains":["agents","developer-tools"],"primitives":["choice","score","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/y0usaf/pi-jev","author":"y0usaf","externalId":"y0usaf/pi-jev","access":"public","publishedAt":"2026-09-16","capturedAt":"2026-09-18"}],"metric":{"name":"Output-judge fixture requests","value":203,"unit":"requests","conditions":"Author-reported smoke calibration across 53 fixtures, not a representative safety benchmark.","sourceUrl":"https://github.com/y0usaf/pi-jev","independentlyVerified":false},"modelVersion":"jev-latest","limitations":["The smoke calibration is small and does not establish reliable safety performance.","Failures fail open, and tool arguments or output excerpts are sent to the TypeSafe API."],"repoUrl":"https://github.com/y0usaf/pi-jev","permissionStatus":"public-source","publishedAt":"2026-09-16","lastVerifiedAt":"2026-09-18"},{"slug":"remove-ad-like-dom-elements","title":"Remove ad-like DOM elements in Chrome","summary":"A Chrome extension finds ad-shaped DOM candidates and asks Jev whether each candidate is a paid advertisement before code removes it.","whatJevDoes":"Jev answers one Noul question per compact candidate description in a batch. Candidate discovery, page-size guards, thresholding, animation, removal, and mutation observation are ordinary extension code. The project is explicitly a demonstration rather than a replacement for a security or privacy-focused ad blocker.","integrationPattern":"Heuristic DOM candidates to batched advertisement probabilities; browser code thresholds and removes matching nodes.","jobs":["classify","control"],"domains":["automation"],"primitives":["noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/realZachi/typesafe-adblock","author":"realZachi","externalId":"realZachi/typesafe-adblock","access":"public","publishedAt":"2026-09-17","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["It can miss ads or remove non-ad content and does not block tracking, malware, or video advertising.","Only candidates found by deterministic heuristics are sent for judgment, so hidden or unlabeled ads can be missed."],"repoUrl":"https://github.com/realZachi/typesafe-adblock","permissionStatus":"public-source","publishedAt":"2026-09-17","lastVerifiedAt":"2026-09-18"},{"slug":"turn-home-state-into-automation-signals","title":"Turn Home Assistant state into automation signals","summary":"A Home Assistant integration exposes Jev probabilities, choices, and scores as entities and action responses that automations can use.","whatJevDoes":"Jev evaluates selected entities, devices, areas, floors, labels, or templates and returns typed answers. The integration handles configuration, update triggers, batching, daily token budgets, thresholds, and entity creation. Automations can then branch on the resulting values without parsing model prose.","integrationPattern":"Home state to typed Jev questions; the integration maps answers to sensors and automation response variables.","jobs":["classify","score","route"],"domains":["home","automation"],"primitives":["choice","score","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/AboveColin/HA-Jev","author":"AboveColin","externalId":"AboveColin/HA-Jev","access":"public","publishedAt":"2026-09-17","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["The author advises against using the integration for safety-critical locks, heaters, or smoke alarms.","Answers contain no reasoning and published confidence values require local calibration."],"repoUrl":"https://github.com/AboveColin/HA-Jev","permissionStatus":"public-source","publishedAt":"2026-09-17","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"prune-stale-agent-tool-history","title":"Prune stale coding-agent tool history","summary":"A Pi extension uses Jev to decide which old tool calls and results still matter while keeping conversation text verbatim.","whatJevDoes":"For each eligible tool call, Jev estimates whether the call and its result should remain in context. Code turns those probabilities into monotonic keep, truncate-result, or drop-call decisions, pins recent messages, records an append-only ledger, and falls back to Pi's built-in summary compaction when pruning is insufficient.","integrationPattern":"Compact tool-history state to paired Noul judgments; code preserves order and applies reversible context filtering.","jobs":["classify","control"],"domains":["agents","developer-tools"],"primitives":["noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/joelhooks/pi-fast-jev-compaction","author":"joelhooks","externalId":"joelhooks/pi-fast-jev-compaction","access":"public","publishedAt":"2026-09-18","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["Tool-call pairing depends on stable Pi identifiers and truncated results can lose attached images.","Model judgments are not semantic proofs, and the extension does not create summaries or durable memory."],"repoUrl":"https://github.com/joelhooks/pi-fast-jev-compaction","permissionStatus":"public-source","publishedAt":"2026-09-18","lastVerifiedAt":"2026-09-18"},{"slug":"expose-typed-judgments-to-mcp-agents","title":"Expose typed Jev judgments to MCP agents","summary":"An MCP server gives compatible agents tools for classification, scoring, checking, matching, screening, and custom typed Jev questions.","whatJevDoes":"Jev provides the bounded Choice, Score, and Noul answers behind six MCP tools. The server validates requests, applies configurable budgets and retry policy, adds abstention to matching, and returns a stable structured envelope. Agent clients receive typed values rather than prose, while consequential authorization decisions remain outside this screening layer.","integrationPattern":"MCP tool calls become validated Jev requests; the server returns stable structured results and keeps policy in code.","jobs":["classify","score","route"],"domains":["agents","developer-tools"],"primitives":["choice","score","noul"],"sourceType":"community","deploymentStatus":"reproducible-repo","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"github","url":"https://github.com/blakestone-x/jev-mcp","author":"blakestone-x","externalId":"blakestone-x/jev-mcp","access":"public","publishedAt":"2026-09-17","capturedAt":"2026-09-18"}],"modelVersion":"jev-latest","limitations":["Screening is not a security or authorization boundary, and confidence is not correctness.","Published measurements come from one early-access key and a limited field-service data set."],"repoUrl":"https://github.com/blakestone-x/jev-mcp","permissionStatus":"public-source","publishedAt":"2026-09-17","lastVerifiedAt":"2026-09-18"},{"slug":"route-smart-home-assistant-requests","title":"Route smart-home assistant requests","summary":"TypeSafe's smart-home demo evaluates a request against many typed questions in parallel, then lets code use only the answers relevant to that request.","whatJevDoes":"Jev classifies the request category, target domain, device type, and requested device action in one speculative batch. It also checks whether the request contains multiple actions. Code ignores irrelevant speculative answers, splits compound requests through a separate language model, and sends conversational requests to that model instead of treating Jev as a text generator.","integrationPattern":"One request becomes parallel Choice and Noul judgments; deterministic code selects the applicable branch and owns device actions.","jobs":["classify","route","control"],"domains":["home","automation"],"primitives":["choice","noul"],"sourceType":"official","deploymentStatus":"official-example","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"demo","url":"https://docs.typesafe.ai/demos/smart-home","author":"TypeSafe","externalId":"demos/smart-home","access":"public","capturedAt":"2026-09-18"}],"limitations":["The official page describes the demo architecture but says the full source repository will be available at release.","The page publishes no accuracy, latency, device-coverage, or production-reliability benchmark for this assistant."],"permissionStatus":"public-source","publishedAt":"2026-09-18","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"batch-regulatory-briefing-questions","title":"Batch a regulatory briefing into one call","summary":"An official cookbook asks 13 regulatory questions over a pinned GDPR article in one Jev request and compares that batch with 13 separate requests.","whatJevDoes":"Jev answers eight Noul, two Choice, and three Score questions against the same 53,777-character document. The cookbook repeats both batching strategies five times and compares each answer's tracked probability or normalized score. Local code pins and fetches the article, caches calls, calculates variance, tokens, cost, and elapsed time, and renders the comparison.","integrationPattern":"A document-dominated workload batches independent typed judgments so shared state is sent once; code measures and compares the two call strategies.","jobs":["classify","score","verify"],"domains":["automation"],"primitives":["choice","score","noul"],"sourceType":"official","deploymentStatus":"official-example","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"cookbook","url":"https://docs.typesafe.ai/cookbooks/parallel_questions","author":"TypeSafe","externalId":"cookbooks/parallel_questions","access":"public","capturedAt":"2026-09-18"}],"metric":{"name":"Batched-call cost ratio","value":12.2,"unit":"× cheaper","conditions":"TypeSafe-reported five-run comparison of one 13-question call with 13 sequential single-question calls over the same pinned GDPR article using jev-1.12; the cookbook separately reports a 10.0× sequential latency ratio.","sourceUrl":"https://docs.typesafe.ai/cookbooks/parallel_questions","independentlyVerified":false},"modelVersion":"jev-1.12","limitations":["The cost advantage is largest because the long shared document dominates each request; workloads with little shared state will differ.","The reported speed ratio sums sequential single-call latencies, and concurrent requests would reduce that latency gap without reducing repeated input tokens."],"permissionStatus":"public-source","publishedAt":"2026-09-18","lastVerifiedAt":"2026-09-18","featured":true},{"slug":"verify-llm-citations-against-source","title":"Check LLM citations against source context","summary":"An official cookbook combines exact quote matching with a Jev relation judgment to label citations verified, unsupported, contradicted, or fabricated.","whatJevDoes":"For each quote that exists in the source, Jev receives the claim and surrounding RFC section and chooses whether the section supports the claim, contradicts it, or says nothing about it. Code normalizes and locates quotes first, marks missing quotes as fabricated without a model call, maps the relation to a verdict, and routes low-confidence answers to human review.","integrationPattern":"Deterministic source matching gates one bounded relation Choice; confidence thresholds decide whether code accepts or reviews the verdict.","jobs":["verify","classify","route"],"domains":["agents","developer-tools"],"primitives":["choice"],"sourceType":"official","deploymentStatus":"official-example","verificationStatus":"artifact-verified","evidenceLevel":"E2","sources":[{"kind":"cookbook","url":"https://docs.typesafe.ai/cookbooks/citation_check","author":"TypeSafe","externalId":"cookbooks/citation_check","access":"public","capturedAt":"2026-09-18"}],"modelVersion":"jev-1.12","limitations":["The published fixture contains eight citations, including four deliberately edited failures, so it is not a representative accuracy benchmark.","Exact matching can label a shortened or lightly paraphrased quote as fabricated, and confidence thresholds require calibration on the target corpus."],"permissionStatus":"public-source","publishedAt":"2026-09-18","lastVerifiedAt":"2026-09-18","featured":true}]}