Execution 0.1.11 fixes release of an HTTP owner that has never attempted Runtime start, while retaining cleanup gates for started or uncertain sessions. Release scope and validation. Amadeus extension 0.0.9 introduces clearer, user-facing descriptions; extension release notes. Store approval is separate from release availability. Affected-installation recovery. Learning stays paused; see archive notes.
English · 简体中文
Execution 0.1.11 — verified execution, updates under your control.
Your image-capable Agent reads the target and result.
Runtime keeps original evidence, input gates and action receipts.
Download execution 0.1.11 · Demos · Quick start · Architecture · Evaluation
Retained API-route recovery: After a target is rejected by both model checks before input, Main can inspect the bound screenshot, explicitly review it and separately continue the same job. Settled text-input review also remains available without retyping. Acceptance and limits →
Agent Review Instant is a Windows computer-use runtime for MCP-compatible agents. Give it an explicit linear plan to navigate an application, click targets, replace text, press keys or scroll. It executes supported steps in one session and checks their results, reducing the need to return to the main agent after every successful action.
Choose Use the current Agent to let your connected, image-capable agent locate targets and check results, or Use an external model API for Runtime-managed visual checks. Neither requires local model weights or a GPU. Local models and delegated vision remain advanced options.
Turn 3,000 synthetic orders into a checked sales report: clean the data, calculate gross profit, build a monthly PivotTable, rank stores and create charts. This English walkthrough shows real desktop actions, with planning, review and corrections by the main Agent.
Watch in English · 3:05 · 4K video · 中文演示
The preview uses selected excerpts at their original playback speed. The full video is an edited walkthrough, including a continuous 20-second 1× PivotTable segment; its duration is not an end-to-end speed benchmark.
Current-agent mode (introduced in 0.1.7): --current-agent provides a persistent basic-click loop without a separate Runtime vision or Decision API. The hosting agent's service and quota still apply. Text input, keys, scrolling, task plans and administrator mode are outside this adapter's supported set. Setup and capability matrix.
- Release a proven never-started HTTP owner: an initialized owner may be deleted with
cleanup_verified=falseonly when lifecycle evidence proves no start attempt, attached resources or pending work. An initial-looking status alone is insufficient. - Keep existing cleanup gates: the
cleanup_verified=truerelease path remains available. Any attempted start, active or unknown work, attached resources or unverified cleanup still requires the original cleanup path. - Connector 0.0.8: recommends Runtime 0.1.11 with unchanged connection settings, permissions, model policy and operator-only recovery. Exact regression, package and installed results belong to the release validation record; no new speed or accuracy result is claimed.
These 0.1.4 demos are compressed previews of the published videos, with lower resolution/frame rate and no playback acceleration. Browser and Notepad clips trim the initial wait; GIF length is not a performance measurement. Full videos and actual plan timings →
Google URL → search java → open the first ordinary result. Full MP4
Edit a note → File → Save. Full MP4
Navigate A → nested → Up → root → B → bundle.zip. Full MP4
| Capability | What it means in practice |
|---|---|
| Keep a task moving | Run multi-step tasks continuously in the same session—for example, open a search engine, enter a query and follow a result. |
| Check the step that just happened | Ground targets from screenshots. Use cropped, annotated evidence to check a target or the requested result with Decision. |
| Recover with the original agent | If Decision does not accept, the standard Luna vision model reviews the same evidence once. Unresolved cases return with evidence; after reviewing it, the original agent can explicitly continue without replaying completed actions. |
You need Windows x64, uv and Python: 3.12 recommended for API/Agent use (existing 3.11 environments remain supported), or 3.11 for local models. You also need an MCP-compatible agent. The installer includes source and setup scripts; first-time setup installs dependencies. Keys and model weights are not bundled.
- Download and install execution 0.1.11. See the release record for downloads, checksums and validation. Versions 0.1.7 and earlier require manual installation; affected 0.1.8/0.1.9 sessions need the settlement procedure first. Old evidence may remain insufficient; recovery of every session is not guaranteed.
- Choose a mode below. For external Luna + Decision, copy the vision profile and Decision profile, then make
OPENAI_API_KEYavailable to the MCP process. For current-agent clicks, use the 0.1.7 entry and declare the client's real image/vision capabilities. - Run setup and connect your agent. Follow the setup commands, merge the generated
agent-review-instantMCP entry into your client, and reconnect. Keep other server entries intact.
| Main mode | Who locates and checks? | Setup |
|---|---|---|
| Use the current Agent | The connected main agent; no separate Runtime vision/Decision API. Uses the hosting agent's service quota. | 0.1.7 entry and supported click loop |
| Use an external model API | Runtime calls configured vision and Decision services. Provider usage charges apply. | API profiles and credentials |
Advanced: local VISTA-4B requires weights and local dependencies; delegated vision uses a selected, reusable visual-agent session.
Default visual input sequences require Decision in auto mode. Without Decision, basic actions and explicit legacy/native routes remain available with main-agent review; the runtime does not silently switch routes. All routes require the local Windows runtime. API usage is billed by the provider.
Provider configuration in 0.1.6: Decision and secondary review each support their own endpoint, model and key environment variable. Defaults remain OpenAI Decisions + ordinary Luna; replacements require Decisions/Responses-compatible services. These settings first shipped in the 2026-10-15 rebuilt 0.1.3 installer. Configuration and compatibility →
Amadeus/Pi: Let Amadeus handle repetitive Windows desktop tasks through Windows Desktop Automation 0.0.9. See what it does and how to set it up. Install and start Runtime separately; 0.1.11 is recommended. Version 0.0.8 is already in the official store; 0.0.9 requires its own catalog review and can be imported manually from the release. The historical 0.0.2 stdio guide remains archived.
Old wheels and download URLs retain their original bytes. Extension 0.0.9 updates presentation and documentation; connection keys, permissions and the Runtime recommendation stay unchanged. It does not establish current-agent mode in Amadeus. See the current release scope, historical 0.0.5 qualification and Runtime update instructions. Disabling or uninstalling the connector does not stop or uninstall the external Runtime.
0.1.7 current-agent source acceptance: 20 live 7-Zip steps on one connection and separate refusal controls passed; see the original-image, receipt and cancellation record. The separate installed five-click Calculator workflow, cancellation controls and cleanup also passed; see the release record. This does not establish a performance gain, cost-saving percentage or general success rate.
0.1.5 refresh — build 20261017-01: adaptive observation after Enter and next-step read-only grounding alongside current-result verification. The refreshed installer completed the following tasks on the same frozen payload:
| Installed task | Round 1 | Round 2 | API requests, both rounds | Reported tokens, both rounds |
|---|---|---|---|---|
| Notepad: edit → File → Save | 20.744 s | 18.191 s | 21 | 37,482 |
| Edge: Bing → search → Wikipedia article | 27.043 s | 27.430 s | 27 | 42,111 |
| 7-Zip: folder / prepared-ZIP navigation | 25.918 s | 24.324 s | 18 | 47,585 |
Six plans / 18 steps completed automatically on their first attempt. Main checked Notepad and Edge; independent Sol checked 7-Zip. Original screenshots, exact Notepad disk contents and owned-window/host cleanup passed. Each run above is a complete three-step plan including visual verification; launch, between-round inspection and cleanup are excluded from its time. 7-Zip compression/extraction was not tested.
The installed runs used 66 API requests: 18 grounding + 46 Decision + 2 ordinary review; 127,178 reported tokens, with no missing usage. Consumed prefetch copies are counted once; Main/worker subscription usage is excluded. These scoped passes do not establish a general success rate.
Measured action intervals, source A → B: 5.803→5.597 s (Notepad), 7.805→7.254 s (7-Zip), 10.786→9.280 s (Edge). These are medians from the first physical mouse-down of one logical step to the next, four matched intervals per application/arm. A already includes the earlier ordinary-click/SSL changes; B adds Enter observation and overlap. This is not a direct old-installer/new-installer comparison. One Edge B plan was slower than A; model/network/cache variation was uncontrolled. The four-second interval objective is not yet reached. Models, prompts and acceptance rules are unchanged.
Verification: 395 affected checks passed; after race fixes, 82 overlapping checks passed (not added as unique tests). All 659 installed payload hashes matched before/after acceptance. A new isolated Python 3.11 environment installed 45 locked API-route dependencies and passed functional-entrypoint checks. Full measurements, source-comparison usage and limits → · Dated installer and checksums →
Original 0.1.5 build, installed Amadeus/Pi route: two normal synthetic Notepad tasks completed six steps, exact disk checks and cleanup; no Main semantic override was needed, and outer self-corrections remain recorded. Final source regression: 135 passed. The final installer independently completed a three-step Notepad plan on its first attempt in 27.70 s / 10 API requests, with exact disk output and cleanup. This is one scoped sample; total token usage is unknown. Scope and limitations → · Release and checksums →
Historical 0.1.4 installed tasks, beyond the browser: fresh Notepad and 7-Zip tasks used the maintained visual execution route.
| Task | Observed result | Plan time |
|---|---|---|
| Notepad, first round: edit → File → Save | All 3 steps completed automatically; saved UTF-8 alpha content matched exactly | 30.621 s |
| Notepad, same-window second round | All 3 steps completed automatically; saved UTF-8 beta content matched exactly | 22.873 s |
| 7-Zip: A → nested → Up → root → B → bundle.zip | All 6 steps completed automatically; the archive's package folder was visible | 69.515 s |
All 12 installed steps completed without Main review; one ordinary review resolved automatically. Owned-window closure and host cleanup passed. Independent Sol then completed fresh 3-step Notepad and 3-step 7-Zip tasks, with disk/image checks and cleanup. The 7-Zip scope is directory/prepared-ZIP navigation; compression and extraction were not tested.
Historical 0.1.4 recorded demos: Notepad, full window · Browser · 7-Zip. The new Notepad/Browser clips skip the initial wait, retaining about one second before first input; later actions and waits play at 1× speed, with no narration. Clip lengths are 19.8 s / 27.2 s; the original complete three-step plans took 28.0 s / 31.4 s. Recording scope and edits →
Separate historical 0.1.4 source02 evidence covers a 12-step plan plus 4-step continuation. Nonempty linear plans have no fixed step-count cap; budgets and evidence checks remain. Earlier source false rejection and public-client receipt-cap failures remain recorded as failures. Current results, exact timings and failure history → · Long-plan regression →
Historical 0.1.3 browser task: the installed candidate completed Bing → Rust collections documentation twice in the same window/session, in 30.809 s / 27.354 s, using 41,097 API tokens, without Main intervention or ordinary Luna review. These are historical task-specific times, not a controlled speed comparison; token totals exclude Main/worker subscription usage. Timing and recovery evidence → · Operation breakdown →
Historical 0.1.3 prompt evaluation used 39 real UI screenshots across 30 scenario families. On the held-out set, three-way label agreement rose from 92.5% to 95.0%, with 1.64% fewer Decision input tokens. Pass/nonpass accuracy stayed at 95%; the two-stage result stayed at 149/160, including 7 false accepts and 1 false rejection with the adopted prompts. Labels were AI-reviewed: these are not general GUI success rates. Method, split and failure cases →
User task → Main-agent plan → Ground → Check → Act → Verify → Continue
In current-agent mode, the connected Agent reads the original screenshot, grounds the target and judges the after image. In the configured API route, the vision model supplies coordinates and OpenAI's Decisions API checks the target or result using Choice. A valid pass continues the plan; a nonpass gets one ordinary Luna review before returning to the original main agent if unresolved. Decision does not replace the planner.
The runtime binds evidence to the request, window and screenshot, and validates coordinates and freshness. Default execution uses the visual route; UIA remains available through explicit native routes. Keep one persistent session and one desktop input owner.
In 0.1.3: shared observations, preloaded plans and reused HTTP connections reduce repeated preparation. The installed two-round task avoided 10 redundant outer captures. Evidence-bound continuation lets the main agent review a completed step without repeating its action.
Historical 0.1.5 — typed-input Main review: after authenticated dual-model nonpass on settled typed input, inspect the bound image and explicitly continue the same command without retyping. Remaining submit actions retain their fresh-target and verification gates. Earlier live04 recovery used explicit TEST injection on an unreleased payload; it is not a natural-error or performance result. Contract → · Acceptance and limits →
Detailed execution flow → · Task-plan contract → · Choice contract →
-
Current-agent adapter: basic clicks only. Text input, keys, scrolling, task plans, automatic cross-window work, restart continuation and administrator/UAC are unsupported by this adapter. Existing API/legacy capabilities keep their own contracts.
-
Since 0.1.4, linear plans: nonempty plans have no fixed step-count cap. Unique step IDs, per-step verification, context-size limits, evidence checks, cancellation and API/time budgets remain. Existing 0.1.3 installations retain their build-time contract until upgraded. Source checks
-
Models can reject correct evidence or accept incorrect evidence. A primary false pass skips secondary review; this remains a known limitation. Arbitrary unattended tasks are not guaranteed.
-
The installer is unsigned. API dependencies were checked in an isolated Python 3.12 environment; new-machine setup and normal OS registry/shortcut integration were not newly verified. Pi full receipts remain verbose; this release makes no token-saving claim. Release scope and history
Each release's current downloads, accumulated by release date as of 2026-11-09; refreshed daily. Includes installers, updates, extensions and repeat downloads. Per-release data.
ISC License. Dependencies and model weights retain their own licenses. The banner illustrates the workflow; it is not a recorded run.
The benchmark backup is available in a separate public repository. The frozen suite contains 39 original screenshots, 30 interface families and 254 labelled cases. Dataset images, annotations and raw evaluation records are not bundled with the execution installer or downloaded during setup. Clone separately when needed.
Release history
- Settle proven pre-dispatch refusals: original durable processing evidence distinguishes a rejected control from unknown input; an error string or missing record alone never permits maintenance.
- Observe without replay: bind read-only observation to its original capture and revised review packet, then settle the completed original task. Unknown, interrupted or altered evidence remains blocked.
- Connector 0.0.7: recommends the tested Runtime combination with unchanged settings and permissions. Advanced
recover/resume_pre_dispatchsettlement remains unqualified. Scope and validation.
- Settle the original plan: maintenance validates the frozen plan, runner and input receipts for completed or proven-cancelled task plans. Active or unknown input and incomplete evidence still block installation.
- Upgrade affected 0.1.8 installations: run the verified new release's maintenance script against the old installation, then use the normal confirmed update path. Do not delete task registrations. Recovery procedure.
- Amadeus connector 0.0.6: recommends Runtime 0.1.9 with unchanged connection settings, permissions and same-session Main review/continue behavior. Release scope and validation. No new speed, accuracy or token-saving claim.
- Check automatically, install after confirmation: launchers check stable releases at most daily, in the background; checks can be disabled. Download/verify with
update_runtime.ps1 -Action stage, then explicitly confirm the exact target with-Action apply. - Keep tasks intact: installation-wide locks and original-session cleanup evidence prevent maintenance while clients, hosts, model workers or unresolved input remain. The updater never closes a task to install.
- Restore a verified version: retain backups and a persistent journal, run a no-input installation health check, and roll back on failure. Interrupted updates have an external recovery entry; unknown states retain backups and block further action. Compatible patches keep the Amadeus wheel unchanged. Commands and limits · Release record.
- Original-image, original-command loop: prepare → inspect original image → resolve → continue → inspect after image → verify. Keep one connection and one input owner; pending reads keep the original IDs and never replay unknown input.
- Source acceptance: 20 live 7-Zip steps passed on one connection. Wrong-hash review, uncertain judgment and input after completion were blocked. Cancellation while awaiting grounding reached real cancelled state with no input, and cleanup passed. No independent vision/Decision profile or provider key was used. Live evidence. Installed 0.1.7 separately passed Calculator C → 1 → + → 2 → =, showing 3, with per-frame Agent review; post-action and awaiting-grounding cancellation controls and cleanup passed. Installed evidence and release.
- Update boundary: a compatible HTTP contract does not require reinstalling an unchanged wheel for each Runtime patch. Runtime and the Amadeus extension have independent release lifecycles.
- Amadeus over HTTP MCP: keep the existing stdio route or use the new authenticated, loopback-only service. The optional 0.0.4 extension connects Pi and supplies its execution Skill; install/start Runtime separately. Install the extension.
- Main can recover before a click: Decision and ordinary Luna nonpasses return the original candidate and bound before image for explicit Main review and separate continuation. The same job continues; completed input is not replayed. Existing input, freshness and cancellation checks remain.
- Queue recovery: temporary command-queue lock contention no longer becomes a permanent task-binding failure. Decision/Luna thresholds, prompts and provider defaults are unchanged.
- Scoped acceptance: a five-step direct Runtime workflow and a four-step native Amadeus/Pi workflow completed after explicit TEST double refusal and actual Main image review. Python 3.11 and 3.12 each passed 188 affected regression checks. These are functional recovery results, not new speed, accuracy or token-saving claims. Evidence and known limits.