User repository on GitHub
冻结计划的自动化中间件（TypeScript / Node ≥ 22.6）：LLM 只规划一次，执行由本地规则与决策模型逐步完成，演示一次后既可端侧执行。
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Integrations & Adapters
AmberFlow
A frozen-plan browser automation middleware for TypeScript and Node 22.6+: an LLM plans once, local rules and decision models execute step by step, and after one demonstrated run it can repeat the plan offline.
- Stars
- 0
- Language
- TypeScript
- License
- APACHE-2.0
- Reading time
- 2 min
- Published by
- Community
- Author
- Johnly1986
- Last verified
- Oct 2026
Works forJev for Realtime control
Used inJev for Software & DevTools
- Languages
- TypeScript
- Platforms
- PlaywrightCLI
- Capabilities
- Open sourceRuns locally
Highlights
- The LLM plans once; execution is handled step by step by local rules and decision models, so runs become reproducible after one demonstration.
- npm start runs a real plan offline on the repository's own page with zero configuration; npm run doctor checks Node, Chromium, models and providers.
- Decision providers include laya-onnx, scripted-stub, none and a first-candidate baseline described as the line any new model must beat.
- Ships as a DeepSeek Harness plugin with three tools: amberflow_apps, amberflow_run with dry, ablate and slots, and amberflow_learn for recording.
- Plan checkpoints go to .task_plan.json in the repo root unless you pass --plan; Apache-2.0, created 2026-09-30, pushed 2026-10-13.
Quickstart
bash
npm ci && npx playwright install chromium
npm startWatch out
README reviewed 2026-10-17; the repository is documented mainly in Chinese and the digest excerpt stops inside the DSH plugin notes, so the full setup and any benchmark numbers are not captured. The jev and systemone topics are the clearest decision-model evidence in the digest.
Questions about this entry
Short answers drawn from the linked project or page.
A browser automation middleware for TypeScript and Node 22.6 or newer. An LLM plans once, then local rules and decision models execute the plan step by step; after a demonstrated run it can execute on-device without replanning.
The repository is tagged jev and systemone and describes execution by local rules and decision models; the captured excerpt names the Laya ONNX provider and a first-candidate baseline, so the exact Jev usage is not visible in the digest.
Run npm ci && npx playwright install chromium, then npm start, which runs a real plan offline on the repository's own page with zero configuration. npm run doctor checks the environment and prints the command for anything missing.
Yes. The README says steps whose targets the plan names are solved by rules, and steps that genuinely need a model state exactly what is missing instead of guessing or throwing, so weights are only needed for those fallback steps.
With npm run apps -- --app ctrip-case --provider coin-flip you can run against laya-onnx, scripted-stub, none or the baseline provider; npm run providers:baseline runs the comparison bench any new provider must beat.
Three tools: amberflow_apps lists apps, amberflow_run runs one with dry, ablate and slots options, defaulting to the real model, and amberflow_learn records and generalises new plans. Install it with dsh plugin add, then restart DSH after a dev link.
The README is mainly Chinese; the commands are readable as-is but the prose is not in English. The project is Apache-2.0 licensed, and GitHub metadata shows creation 2026-09-30 with the last push 2026-10-13.
More Jev resources like this
A local stdio MCP server from bgivenb that runs whole browser and macOS goals for your agent: your agent supplies the goal, exact values, and completion conditions, and Flick loops observe, decide, and act with TypeSafe Jev. TypeScript core with Playwright and Swift drivers.
Jev Agents & MCP#community#github#mcp
Community
bgivenb/flick-computer-useTypeScript
bgivenb/flick-computer-useTypeScriptA host-agnostic routing layer that sits between a coding harness and its models, choosing capacity per round from trajectory evidence like tool calls, failures and context pressure, running as a user-scoped daemon and local OpenAI-compatible proxy.
Jev Integrations & Adapters#community#catalog-import#logicrw
Community
vizuh/sabiTypeScript
vizuh/sabiTypeScriptAn auto-mode plugin for the Pi coding agent that classifies shell commands before they run: built-in rules decide most calls, and Jev (hosted) or Kev and Laya (local, CPU, offline) route the rest to run, confirm, or block.
Jev Integrations & Adapters#community#github#pi
Browser QA where Playwright drives and films, and Jev judges: natural-language assertions, self-healing locators and goal-driven click loops, with a JSON-flow CLI for agents, a local dashboard and an agent skill.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
jonymusky/jev-browser-qaTypeScript
jonymusky/jev-browser-qaTypeScriptA Claude Code mod that sends each prompt to Jev through OpenRouter and turns the returned 0-3 score into a reasoning effort level, from low to xhigh. Shadow mode is on by default, so it only logs what it would set until you turn it off.
Jev Integrations & Adapters#community#github#claude-code
Community
gebeer/jev-effortTypeScript
gebeer/jev-effortTypeScriptA bridge and Claude Code plugin that runs scripts of taps, typing and swipes on iOS simulators or Android devices and has Jev judge each checkpoint's screen text, returning passed, failed or inconclusive with screenshots and a log. Saved scripts rerun from the terminal with no Claude in the loop.
Jev Integrations & Adapters#community#github#ios
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.