mend.

MEND AI · MODEL v1.0.0

Watch Mend AI learn.
One verified fix at a time.

Mend AI is the small model inside every Mend server. When a screen in your app gets stuck, it chooses one of the recoveries you approved, applies it, and learns from whether the browser confirms the next screen really appeared.

When it is not sure, it asks Claude. Claude’s checked answer can only make Mend AI more careful, never take over. Confidence comes from browser-verified results alone.

Try the free demo See the road to a local LLM
How Mend AI v1 is wired Five kinds of transition failure on the left connect through Mend AI's learned memory in the centre to five developer-approved recoveries on the right. v1 Hidden screenMissing screen Slow transitionCancelled animation Slow animation Reveal screenRetry navigation Reduce motionRedirect to fallback Ask the developer
How Mend AI v1 is wired: what it sees on the left, what it can do on the right, and its learned memory in the middle. An illustration, not live data.

LIVE LEARNING FEED

Live activity starts with the next server release on October 11, 2026.

From then on, this page will show Mend AI learning from real, browser-verified fixes as they happen: which kinds of failures it learned from, which recovery worked, how many fixes it now makes on its own, and how many corrections it took from Claude.

  • Anonymized counts by category only: never names, labels, URLs, element IDs or which customer it came from.
  • Small categories are hidden so no single customer’s activity can be picked out.
  • Any workspace can opt out, and turning learning off in a workspace stops it learning there.

THE LOOP

Train → fix → iterate.

  1. 1 · TrainLearns from verified fixes

    Every recovery the browser confirms (or rejects) updates what Mend AI knows about that kind of failure, on that server.

  2. 2 · FixClaude corrects, cautiously

    When Mend AI is unsure or a recovery fails, it asks Claude. Only the action Claude chose is counted, never Claude’s text, and it can only make Mend AI more careful.

  3. 3 · IterateImproves every night

    A nightly job tests Mend AI on synthetic transitions and proposes better rules. Each change is reviewed before release and becomes a new version.

BENCHMARK · v1.0.0

Where v1 starts.

Measured on Mend’s synthetic test suite of transition failures, not on customer data. These are the numbers each new version has to beat.

Fixed and verified
63.9%

of synthetic failures recovered with a browser-verified result.

Asked Claude
63.4

times per 100 transitions. The goal is to bring this down as Mend AI learns.

Wrong recoveries
16.1%

tried a recovery that did not verify. Mend reports these to the developer instead of hiding them.

VERSION HISTORY

Mend AI versions.

  1. v1.0.0October 10, 2026

    First release. A learned micro model: outcome counts and safe rules that choose among developer-approved recoveries, with Claude for the cases it is unsure about.

  2. v1.xNext

    Nightly train → fix → iterate improvements, each reviewed and released as a new version with its benchmark results listed here.

ROADMAP

From micro model to flagship local LLM.

  1. Nowv1 · Micro model

    Runs on every Mend server. Learns from verified fixes. Asks Claude only when unsure.

  2. Nextv1.x · Sharper micro model

    More fixes handled locally, fewer Claude calls, and faster recoveries as the nightly loop and real verified outcomes improve it.

  3. Plannedv2 · Flagship local LLM

    A small language model that runs on your own server or device, trained on browser-verified outcomes and developers’ own registered fixes — never on Claude’s text. Most fixes would then need no cloud AI at all.

Honest status: today’s Mend AI v1 is a learned micro model, not yet a language model. v2 is planned and not built yet.