fastpilot

io.github.autobot-lab由 GitHub 帳號 autobot-lab 發布v0.5.1更新於 Nov 22, 2026

AI-driven QA for web, Android and iOS: drive Chrome, emulators and simulators.

已驗證STDIO僅桌面瀏覽器自動化開發者工具

概覽

AI 產生的概覽

面向 QA 的自動化工具,可驅動 Chrome、Android 模擬器與 iOS 模擬器:開啟工作階段、執行 YAML 測試流程、斷言並產生報告。

可以做什麼
  • 啟動、停止與列出瀏覽器或裝置工作階段,也可附加到自己已開啟的 Chrome
  • 執行跨工作階段與平行區塊的 YAML 測試流程,自動等待並回傳每一步的耗時
  • 在不執行的情況下檢查流程語法與規則,並斷言可見性、文字或啟用狀態
  • 擷取工作階段的 JPEG 截圖與精簡文字快照
  • 儲存、載入與刪除登入狀態,讓測試略過登入
  • 結束一次執行並產生 Markdown 與 HTML 報告、步驟日誌與斷言結果,並查閱內建參考文件
適用情境
適合讓助理端到端測試 Web、Android 或 iOS 應用程式並回報通過或失敗的情況,例如回歸測試、發行前檢查,或把描述的情境變成可重複執行的測試。
執行需求
透過 npx 在本機執行(需要 Node.js)。需要 Mac 或 Linux 以及 Google Chrome;Android 模擬器與 iOS 模擬器需額外設定,iPhone 測試需要 Xcode。選用環境變數包括用於加密登入狀態的 FASTPILOT_STATE_KEY 與 FASTPILOT_STATE_EXPIRE_DAYS;測試密碼以名稱引用並透過環境變數提供。
安裝前請注意
它會操作真實的瀏覽器與裝置,可以點擊、輸入與上傳,對線上網站會造成實際影響。付款與刪除帳號類按鈕預設會被攔截,一般刪除操作需要確認,但這些限制可由你自行放寬。儲存的登入狀態包含 cookie 與 localStorage,僅在設定 FASTPILOT_STATE_KEY 時加密。防護規則存放在本機設定檔中,任何以你的身分執行的程式都能修改。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 fastpilot,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

fastpilot

Test your web, Android, and iOS apps in plain language, from one dashboard. You say what to check; AI runs it on your own computer, shows you every step live, and tells you what passed and what failed. Free, open source, no cloud. Payment and delete-account buttons are blocked from the start.

[The fastpilot dashboard: an agent called QA Checkout gets one sentence in Chat and answers with a Save test and schedule card that lists the test steps, a schedule every 2 hours, and the buttons Save and schedule, Save test only and Edit.]

Start in 3 steps

You need a Mac (or Linux), Google Chrome, and Claude Code signed in.

1. Install (once)

bash
npm install -g fastpilot

2. Set up your project (once per project)

bash
cd ~/your-projectfastpilot init

init asks before each change; answer y. At the end it asks "Open the dashboard now?": press Enter.

Using Claude Code? You can also install fastpilot as a plugin (the MCP server and its 24 skills in one step): /plugin marketplace add autobot-lab/fastpilot, then /plugin install fastpilot@fastpilot. Use either the plugin or fastpilot init in a project, not both: both together start two fastpilot servers. With the plugin, a secret such as DEMO_PASS works like this: add its name in fastpilot ui › Settings › Secret, and set its value in the environment you start Claude Code from (export DEMO_PASS=…). The value is never typed into the dashboard or the chat.

Just want to look around first? Run fastpilot demo instead. It makes a demo folder with demo pages and example tests, and opens the dashboard on it.

3. Open the dashboard (every time)

bash
fastpilot ui

Your browser opens the dashboard. Keep the Terminal open while you use it; Ctrl-C closes it.

  • The link works once, for 10 minutes. If it expired, press Ctrl-C and run fastpilot ui again.
  • The dashboard only works on this computer.
  • Open it yourself; do not ask the AI to open it for you.

Everything else happens in the dashboard.

Guide: your first test in the dashboard

1. Check your devices: Device

Shows whether Chrome, Android, and iOS are ready, with a ✓ or the reason why not.

  • Install Chrome and Set up an emulator are one-click Fix buttons. Each asks you first, showing the size and the time it takes.
  • iPhone testing needs a one-time setup with your Xcode (about 5 minutes).
  • Only web? Chrome is enough.
[fastpilot's Device page: two Android emulators installed on this computer, each with a Start button, and the iPhone simulators tests can use.]

2. Pick the AI and store passwords: Settings

  • AI model: pick the provider, then the model. Claude Code is the default. Other providers (OpenAI, OpenRouter, Groq, Ollama) also work, and are marked "not tested yet".
  • Secret: add the names of passwords your tests need, for example TEST_PASS. Tests write them as ${TEST_PASS}, and they always show as ***. Never type a password in Chat.
  • Limits & permissions: see the next section, "Safe from the start".

3. Make an agent: Agents › New agent

An agent is a tester with its own job, devices, instructions, and skills.

  • Fill in three things: what it tests, where (Web, Android emulator, iPhone simulator…), and its instructions. Click a suggested instruction to add it.
  • Skills are ready-made test know-how, grouped by technique: boundary values, decision tables, regression, safe checkout, and more. Suggested skills are already ticked, each with a reason. With Atlassian's MCP server in Claude Code, three Jira skills turn tickets into tests, log a failed run as a bug and post results; each write to Jira waits for your OK (setup).
  • Or simply describe the agent in one sentence in Chat and click Create.

You can have several agents, for example "Checkout" on web and "Login" on Android. Each has its own Chat, Live, Schedules, Instructions, Skills, History and Agent settings.

4. Ask for a test: Chat

Write what you want in plain words:

You want toType in Chat
Check something nowCheck that login on https://yourapp.com goes to the Dashboard
Save a testCreate a test for the checkout flow and save it
Test a field properlyTest the quantity field properly (it picks the technique and shows what it covered)
Run everythingRun a regression (newly failing tests come first)
Before a releaseIs staging ready to ship? (one verdict: SHIP, SHIP WITH FIXES, DO NOT SHIP)
Understand a failureWhy did the checkout test fail?

When a test is ready, a card shows its steps: Save & schedule or Save test only. Nothing is saved before you click. The picture at the top of this page shows such a card.

5. Watch it run: Live

Watch every test step by step, on web, Android, and iOS. Take over if you need to tap something yourself, then hand it back to the AI. Several agents can run at the same time; a phone runs one test at a time and the others wait their turn.

[The Live page: the QA Checkout agent is testing the event tickets page in Chrome; four steps are done and the fifth, waiting for the reservation message, is running, with Enlarge and Take over buttons.]

6. Run it automatically: Schedules

  • Every N hours (with active hours, for example 08:00–20:00), every day, or every week.
  • Run now for any schedule.
  • Release watch: after a release, re-check every few minutes for a few hours and only tell you about new problems.
  • Get notified by Slack webhook, Mac notification, or GitHub issue when something fails.
  • Or just ask in Chat: "run the checkout tests every 2 hours".
[The Schedules page: a test that runs every 2 hours from 08:00 to 20:00 and a release watch that checks every 10 minutes, each with its next run, a passed last run and a Mac notification on failure.]

7. See results: History and Memory

  • History: every run and every Chat conversation, with report, screenshots, and video. Delete one, several, or all, or let them be deleted after 7, 30, or 90 days.
  • Memory: how each test is doing (Reliable, Unstable, Failing, Set aside), plus project notes every agent reads, for example "Staging is https://staging.yourapp.com" or "Login uses a code".

Dashboard map

MenuWhat it's for
AgentsYour testers: Chat, Live, Schedules, Instructions, Skills, History, Agent settings. Also the skill Library
LiveEvery running test, from all agents
SchedulesAll schedules and release watches
MemoryTest health and project notes
HistoryPast runs and conversations
DeviceChrome, Android, iOS readiness and Fix buttons
SettingsAI model, Limits & permissions, Secret

Safe from the start

  • Buttons like pay, buy, checkout, transfer, delete account are always blocked.
  • Plain delete/remove buttons must be confirmed first.
  • The AI can make the rules stricter, never looser. To loosen one, you approve it: Chat shows a red card that quotes what you asked, and allowing Pay also needs you to type "Pay". Text on a tested page can never ask for a change.
  • Every change is kept in ~/.fastpilot/audit.jsonl.

Want to test a checkout for real? Ask in Chat: "Allow the Pay button for this project", or open Settings › Limits & permissions.

These limits protect you from what a tested page tells the AI. They are not a lock against programs on your own computer: anything that runs as you can edit ~/.fastpilot/config.toml.

Prefer the terminal?

Everything also works from Claude Code. Open Claude in your project folder (claude) and say it in plain words, for example "fastpilot, run a regression" or "fastpilot, test the login page". Results, schedules, and history show up in the dashboard too.

In GitHub Actions, use autobot-lab/fastpilot-action@v1: it runs your tests on every pull request and adds the JUnit and HTML reports.

If you get stuck

Open Device in the dashboard, or run:

bash
fastpilot doctor
ProblemTry
The dashboard link doesn't workIt works once, for 10 minutes. Press Ctrl-C, then run fastpilot ui again
Chrome or Android isn't readyDevice › Fix buttons, or fastpilot setup web / fastpilot setup android
iPhone isn't readyfastpilot setup ios (once, with Xcode)
fastpilot doesn't show up in ClaudeRun fastpilot init again
A test stops with "Blocked by limits"That's on purpose. See Settings › Limits & permissions
A saved test can't find its password in fastpilot runExport it first, e.g. export FASTPILOT_VARS=TEST_PASS TEST_PASS=...

Learn more

License: Apache-2.0

來源:README.md,提交 71421b9

工具

10

assert

Deterministic assertion with auto-retry: visible | not_visible | text_equals | contains | enabled. Returns pass, or fail with reason and snapshot.

輸入參數kind必填textsessiontest_idtimeoutselector必填

check_flow

Validate a flow without running it: syntax, unsupported steps, unknown sessions, parallel rules, missing ${VAR}s. Errors carry line numbers.

輸入參數flow必填

config

fastpilot settings (~/.fastpilot/config.toml; environment variables win): list, get, set, unset. Changes apply to the next action, no restart. Guard rails (allowed_domains, confirm_actions, block_actions, vars, state_expire_days, allow_* permissions, upload_dirs, max_steps, test_timeout) can only be made stricter here (a running fastpilot also ignores a looser value written to the file until it restarts); program paths and locations (chrome, claude, dir, ios_runner) only by the user with `fastpilot config set` or the file.

輸入參數keyvalueaction

docs

fastpilot reference docs. Topics: dsl, selectors, parallel, examples, errors, memory, devices, skills (the testing skills: regression, bug-report, test techniques, ...), skill/<name> (one skill to follow).

輸入參數topic

report

Close a test run: writes .fastpilot/runs/<ts>-<test_id>/report.md and a self-contained report.html (steps, timings, diffs, screenshots, recordings) plus steps.jsonl and assertions.json, and, when status is pass, saves the flow as path memory for the next run.

輸入參數notesstatus必填test_id必填quarantine

run_flow

Run a YAML flow (batch of steps, may span sessions and `parallel:` blocks). Auto-waits, stops at the first failure, and returns per-step timing plus a diff (or full snapshot) per touched session. Pass test_id to record evidence for `report`.

輸入參數flow必填sessiontest_idstop_on_fail

screenshot

JPEG screenshot of a session, for visual fallback and evidence.

輸入參數sessionannotate

session

Session lifecycle: start (returns the first snapshot, and the path-memory hint when test_id is given), stop, list, devices, start_device. Web sessions launch their own Chrome or attach to yours (mode: attach).

輸入參數idmodeaction必填devicetargettest_idendpointplatform

snapshot

Compact text snapshot of a session (one node per line, stable refs eN). Rarely needed: run_flow already returns state.

輸入參數filefullcoordsfilterformatsessioninteractive

state

Saved login state (web: cookies + localStorage), so a test can skip the login: save | load | list | delete. Files are owner-only, expire after FASTPILOT_STATE_EXPIRE_DAYS (30) and are encrypted when FASTPILOT_STATE_KEY is set. Same as the saveState / loadState steps.

輸入參數nameaction必填session

版本歷史

1
  1. v0.5.1最新Nov 21, 2026

評分

已滿足適用檢查項的 90%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通25 分
  • 已宣告認證方式10 分

    請在註冊中心條目中宣告憑證請求標頭,或保持端點能夠回應探測,以便判定其認證方式。

  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分