Sandbox
@Salmonbird/evoui-browser

Agent-browser skill for verified browser step reuse

Evoui Browser plugs into agent-browser as a reusable skill for browser tasks. It lets the agent do a task normally the first time, captures evidence, and then replays only the verified mechanical parts when the same site and task match again. The host agent still handles judgment, targets, and next actions.

102 stars0 forksPythonUpdated 1mo ago
Who it's for

Builders who run Codex, Claude Code, or Cursor and want repeat browser work to be reused instead of re-explored.

What it delivers

You can reuse verified browser routines on later runs instead of making your agent rediscover the same pages and controls.

What it does

Selective reuse of browser steps

Learns navigation and routine actions only when there is enough evidence and a clear future use.

Host agent keeps control

Agent-browser still chooses targets, interprets results, and decides what to do next.

Page-match checks before replay

Verifies the current page state before reusing any stored routine.

Safe fallback on mismatch

Returns control to the agent when a page changes, a target is missing, or a postcondition fails.

Visible handling of sensitive steps

Pauses for login, CAPTCHA, permission, and confirmation steps in the browser window.

Local run data and knowledge store

Keeps browser profiles, run evidence, and learned knowledge in `~/.evoui/` by default.

How to get it

  1. 1Install the Skill and its pinned browser runtime
    npx skills add Salmonbird/evoui-browser --skill evoui-browser -g
    npm install -g agent-browser@0.33.1
    agent-browser install
  2. 2Then give the agent a concrete website and goal
    Use $evoui-browser to open <website> and complete <task>.   # Codex
    Use /evoui-browser to open <website> and complete <task>.   # Claude Code / Cursor
  3. 3Prefer assisted setup? Send your agent this prompt
    Install Evoui Browser from https://github.com/Salmonbird/evoui-browser.
    Use the README's Quick start, make it available to the supported host I am using,
    keep agent-browser pinned to 0.33.1, and tell me when I can start a new session
    and invoke the Skill.

README

Evoui Browser — reuse verified browser steps

Less repeated browser work. The agent keeps the judgment.

English · 简体中文

CI status Apache 2.0 license Experimental status

Evoui Browser is an optional Agent Skill for agent-browser. Run a browser task normally once; on later, similar tasks, Evoui can reuse verified navigation and mechanical steps so the agent spends less time exploring and repeating work. The host agent still chooses targets, interprets results, and decides what to do next. The currently supported hosts are Codex, Claude Code, and Cursor.

[!WARNING] Evoui Browser is experimental. Start with non-critical tasks on non-sensitive websites.

See the difference

▶ Watch the 71-second invoice-review demo

The first run explores the site step by step. On the next run, Evoui reuses the matching navigation and review routine while the agent still decides which invoices need attention. That avoids rediscovering the same invoice page and controls on every run.

Evoui does not promise that every task will be faster. It helps when a site and task contain repeatable work that can be verified. A first run may take the same amount of work—or more—because it also gathers evidence for possible reuse.

Quick start

Requirements: macOS or Linux, Python 3.10 or newer, and Node.js 24 or newer with npm. Evoui adds no third-party Python packages. Windows is not currently supported because Evoui uses POSIX fcntl.

Install the Skill and its pinned browser runtime:

npx skills add Salmonbird/evoui-browser --skill evoui-browser -g
npm install -g agent-browser@0.33.1
agent-browser install

On Linux, use agent-browser install --with-deps if browser libraries are missing. The skills CLI may ask which detected agents should receive the Skill; choose the Codex, Claude Code, or Cursor hosts you use, then start a new agent session. Some compatible agents share the CLI's global Skill directory, so Evoui may also appear in other detected agents; only the three hosts named above are supported.

[!NOTE] The third-party skills CLI collects anonymous usage telemetry for confirmed public skills. Set DISABLE_TELEMETRY=1 before the npx command to opt out.

Then give the agent a concrete website and goal:

Use $evoui-browser to open <website> and complete <task>.   # Codex
Use /evoui-browser to open <website> and complete <task>.   # Claude Code / Cursor

Prefer assisted setup? Send your agent this prompt:

Install Evoui Browser from https://github.com/Salmonbird/evoui-browser.
Use the README's Quick start, make it available to the supported host I am using,
keep agent-browser pinned to 0.33.1, and tell me when I can start a new session
and invoke the Skill.

How it helps

  1. Run normally. The agent completes the task with agent-browser while Evoui records evidence from the managed calls.
  2. Learn selectively. Evoui keeps only navigation or routines with a clear future use and enough evidence. A successful task does not have to produce knowledge.
  3. Reuse when the page still matches. On a later task, Evoui verifies fresh page state and reuses only the applicable mechanical steps.

Run normally, learn selectively, and reuse verified steps only when the current page matches

Safety boundaries

  • Evoui replays mechanical browser work, not business judgment.
  • A page mismatch, missing target, failed postcondition, or unknown outcome returns control to the agent.
  • A click, submission, or other mutation that may already have taken effect is not blindly retried.
  • Login, CAPTCHA, permission, and confirmation steps pause for you in the same visible browser window.
  • One-off tasks and operations with a reliable API or dedicated connector should use the simpler direct path.

Local data and privacy

Evoui stores browser profiles, temporary Run evidence, and learned knowledge in ~/.evoui/ by default. Managed evidence may temporarily include command arguments, non-empty stdin, page text, form input, or site-returned data. Normal terminal cleanup deletes the content-bearing Run trace after the task and any selected learning work finish; an interruption or cleanup failure may leave data on disk.

Cleanup reduces retained data. It is not redaction, encryption, credential protection, or a privacy boundary. Debug tracing retains additional data without automatic rotation. See SECURITY.md for the full boundary, and never attach credentials, cookies, browser profiles, or task traces to a public issue.

By default, the selected Session and visible browser window remain open after a task for reuse, including their login state. Ask the agent to close the Session when you no longer want it running.

Update or remove

For a CLI-managed installation, update by running the install command again with the supported hosts you use. For example, to update Codex:

npx skills add Salmonbird/evoui-browser --skill evoui-browser -g -a codex

Use -a claude-code or -a cursor instead, or repeat -a for multiple supported hosts.

Existing clone-and-symlink installations continue to update directly:

git -C ~/src/evoui-browser pull --ff-only

After either update method, start a new agent session.

To remove the global CLI-managed installation from all hosts:

npx skills remove evoui-browser -g

Removing the Skill does not delete ~/.evoui/ or the separately installed agent-browser runtime. Inspect local state before deleting it.

If installation fails, first check agent-browser --version. Use agent-browser doctor only to diagnose a browser or daemon startup problem; it performs active checks and may clean stale daemon state.

Contributing

Small fixes, docs, tests, reproducible cases, and focused pull requests are welcome. The two local validation commands and the few changes worth discussing first are in CONTRIBUTING.md. To inspect or modify Evoui, clone the repository rather than relying only on the CLI-managed installation.

License and attribution

Evoui Browser is licensed under the Apache License 2.0. Third-party notices are in NOTICE. Modified documentation derived from agent-browser retains a separate upstream license copy.

Files in the repo

Repository payload12 top-level entries
  • .github
  • assets
  • skills
  • tests
  • .gitignore
  • AGENTS.md
  • CONTRIBUTING.md
  • LICENSE
  • NOTICE
  • README.md
  • README.zh-CN.md
  • SECURITY.md

Discussion (0)

Ask about usage, or say what you built with it

Sign in to join the discussion.

No comments yet. Be the first to say what this is good for.

More skills

obra/
superpowers

An agentic skills framework & software development methodology that works.

285k
1 add

Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.

117k
1 add
Vincentwei1021/
anything2explainer

Topic in, narrated explainer video out. A Claude Code / Codex skill that turns any topic into a black-canvas motion-graphics explainer video with TTS voiceover, subtitles and a chapter progress bar. Chinese or English; every frame drawn in code with Remotion.

666

Open-source AI job search: scan job portals, evaluate listings into a structured A-H report with a global 1-5 score, tailor your CV, track applications — runs locally in your AI coding CLI (Claude Code, Codex, OpenCode, Antigravity…)

71k