Eric Task Master

English | 简体中文

A durable browser automation task system for AI agents.

Version: 2.1.1

AI agents can reason, plan, and write code, but browser execution is often their weakest link. Built-in agent browsers are convenient for short sessions yet commonly lose login state, task continuity, and recovery context. Thin CDP controllers offer fast low-level access, but leave every Agent to rebuild orchestration, progress tracking, cleanup, and error recovery for each job.

Eric Task Master fills that gap. It is a Playwright-first browser automation runtime that turns web work into durable, isolated, observable jobs. An Agent can start a task, leave it running for hours, reconnect later, inspect evidence, recover from a checkpoint, and verify that the browser closed—without occupying the user's everyday browser or repeatedly spending tokens rediscovering the same execution mechanics.

AI agent · browser automation · Playwright · MCP server · web automation · computer use · workflow automation · browser agent · headless browser · multi-agent automation

What can an Agent build with it?

Task Master is a general execution base, not a site-specific bot. Combined with an Agent, a specialized Skill, or a Task Pack, it can support:

• unattended research, monitoring, and source-backed web collection;

• long-running batch work across hundreds of pages without keeping an Agent turn open;

• authenticated back-office, portal, CMS, CRM, and account workflows through isolated persistent Profiles;

• clean no-login jobs through disposable ephemeral Profiles;

• parallel trusted-local Agents sharing a managed Profile catalog while different live Profiles run independently; same-Profile work queues safely and task records stay attributable;

• recurring operational workflows invoked by an Agent, scheduler, or business system;

• browser QA, content verification, form processing, file upload/download, and evidence capture;

• production Task Packs for lead operations, ecommerce operations, media research, supplier research, customer support, compliance checks, or internal RPA.

Those are composable possibilities, not bundled permissions or claims that every website is supported. Task Master supplies the reliable browser runtime; a specialized Skill or Task Pack supplies the business logic, site policy, parsing, checkpoints, outputs, and completion evidence.

Related MCP server: OpenChrome

Why not just use an Agent browser or a CDP script?

Task Master does not replace Agent reasoning. It gives that reasoning a dependable pair of browser hands.

The three-layer model

1. Task Master runtime — pure Playwright execution, persistent and ephemeral Profiles, queues, durable tasks, progress, recovery, evidence, and cleanup.

2. Owner Console — one fixed local web address manages shared Profiles, connected Agents, live tasks, progress, commands, and human-readable reports. Users sign in directly inside isolated persistent Playwright Profiles.

3. MCP + Skills + Task Packs — Agents receive a compact, high-level task interface while reusable domain capabilities stay independent from the core runtime.

This separation keeps the base universal: improve one execution engine, then define many specialized automation workers above it.

For AI Agents: one fixed bootstrap protocol

If a user gives you this GitHub URL, follow the steps below in order. Do not invent another daemon, port, browser launcher, controller, or retry loop.

The full repository is the runtime source of truth. On a GitHub Release, eric-task-master-vX.Y.Z.zip is the complete runnable project; eric-task-master-skill-vX.Y.Z.zip is only the optional Agent instruction adapter and cannot run by itself. SHA256SUMS verifies both archives. Clone the matching tag or download the complete project archive—do not substitute the Skill-only archive for the runtime.

1. Authenticate to GitHub if the repository is private, then clone the complete repository.

2. Require Node.js 20 or newer.

3. Read skills/eric-task-master/SKILL.md. Install that folder into the host's Skill directory when the host supports Skills; otherwise follow it in place.

4. From the repository root run exactly:

   node scripts/taskmaster.mjs connect --json

5. connect installs the lockfile-pinned runtime and matching Playwright Chromium, safely migrates an idle older Manager, starts the loopback Manager, runs real-browser acceptance, and registers supported local MCP hosts. If an older Manager still has work, migration stops without interrupting it.

6. Follow the returned nextAction. After correcting the named precondition, retry the same command at most once; do not branch into a speculative replacement controller.

7. Open the returned Owner Console link once. It silently establishes a persistent local session; there is no code to type or Agent-binding flow. Bookmark http://127.0.0.1:19946/dashboard for later use.

8. If a host reports registered_pending_restart, ask the user to reload that Agent host once. If it cannot reload during this run, use the scoped CLI with one stable Agent ID for this task only, then use MCP for new tasks after the next restart.

9. Choose one operation path and keep it for the task:

   • if the host loaded the registered MCP server, call taskmaster_status, then taskmaster_profiles_list;

   • if the host reports needs_adapter, or cannot reload a pending registration during this run, use the fixed CLI from the repository root. Keep one stable, distinct Agent ID on every scoped command:

     node scripts/taskmaster.mjs status --agent-id STABLE_ID --agent-name AGENT_NAME --json
     node scripts/taskmaster.mjs profiles list --agent-id STABLE_ID --agent-name AGENT_NAME --json

10. When status and Profile discovery succeed, ask the user what browser task to run. Do not mix MCP and CLI identities during one task.

Copyable request for a new Agent:

Install and start https://github.com/npcworkspace-cmyk/eric-task-master. Clone the full repository, read or install skills/eric-task-master, and run only node scripts/taskmaster.mjs connect --json. Do not invent another controller or port. Use the loaded MCP path when supported; on needs_adapter, follow the Skill's fixed CLI path with one stable, distinct Agent ID. Return the Owner Console link; after status and Profile discovery succeed, ask what task I want to run.

Everyday use

After the first bootstrap, the user should be able to ask naturally:

Use Eric Task Master with an ephemeral Profile and adaptive behavior to research these sites, report progress, save evidence, and close every task window when finished.

The Agent then uses one durable loop: discover a task type, start once with an idempotency key, retain the task ID, wait or reconnect to that ID, inspect diagnostics when attention is required, and accept completion only after evidence and cleanup are verified.

Every task start returns a clickable Owner Console link focused on that task. The first link silently establishes the local Owner cookie; later visits can use the fixed bookmarked address. If the user says “启动任务面板”, use MCP taskmaster_dashboard_open or CLI node scripts/taskmaster.mjs dashboard-open [TASK_ID] --agent-id STABLE_ID --agent-name AGENT_NAME --json, then return the link. Task Master does not automatically open an operating-system browser.

Profiles

• persistent — isolated reusable state for logged-in and recurring work; open it from the Dashboard and sign in directly in its Playwright window;

• ephemeral / 隐身临时 — a clean non-persistent browser for each no-login task, destroyed after cleanup.

New persistent Profiles default to the local stable Chrome channel and fixed human behavior. New ephemeral Profiles default to the project-pinned Chromium and adaptive behavior, with fast, adaptive, or human selectable on the Profile. The engine is immutable and never falls back automatically. Manual persistent-Profile windows are always visible; headless affects task runs only.

Behavior

• fast — an optional speed-first policy for ephemeral Profiles and deterministic, data-heavy work;

• human — bounded mouse, typing, scrolling, and reading cadence;

• adaptive — starts fast and temporarily becomes cautious or human-paced after dynamic-page signals, occlusion, timeout, uncertain navigation, action failure, or rate limiting.

Human pacing is a reliability policy, not a promise to bypass website controls or protect an account from platform enforcement. Behavior is selected on the Profile; task start does not accept a behavior override.

Multi-Agent workbench

• Profiles are shared by all trusted local Agents; there is no meaningless “Profile creator” field. A Profile still has one live lease, so two Agents cannot corrupt the same login state.

• The Owner Console shows which Agents are registered, online, offline, working, or revoked, plus their active tasks, Profiles, and queue depth.

• The task workbench shows what Playwright is doing now, progress, limits, recovery state, and a timeline. The Owner can pause, resume, terminate, revise queued input, or send an ask/modify instruction.

• Commands are durable and revision-checked. An active Agent receives them immediately through its wait/inbox loop; an offline Agent receives them after reconnecting. Task Master does not pretend it can wake a completely closed host process.

• Completed work opens on the Agent's human-readable report. Code, logs, screenshots, artifacts, and diagnostics remain available as secondary evidence instead of becoming the “result.”

Build specialized production workers

Keep site and business logic outside the core:

node scripts/taskmaster.mjs task-packs scaffold ./my-pack --name my-pack --recipe paginated-list --json
node scripts/taskmaster.mjs task-packs validate ./my-pack --json
node scripts/taskmaster.mjs task-packs install ./my-pack --json

A Task Pack defines reusable task types. Five production scaffolds—single page, paginated list, list/detail, resumable batch, and form workflow—remove most boilerplate, while preflight validates modules before registration. A specialized Skill teaches the Agent when to use them, how to interpret results, and which platform rules apply. Neither should rebuild Manager startup, browser lifecycle, task tracking, diagnostics, or cleanup. Publish a new task-type name, verify it, then deprecate the old type with a replacement pointer; do not overwrite divergent production logic in place.

Host integration

The browser runtime remains usable through its fixed, Agent-scoped CLI when a host-specific MCP adapter is unavailable. Every scoped CLI command requires a stable, distinct --agent-id for each independent Agent; all Agents share the Profile catalog, while reusing an ID intentionally shares that identity's task ledger and Owner inbox. See docs/MCP-HOSTS.md for the complete non-MCP command path and its trusted-local boundary.

Verification and shutdown

Run the complete local gate:

npm run check

It covers static boundaries, unit/integration/security tests, real Chromium acceptance, and concurrent/fault/restart workloads. GitHub CI verifies Windows, macOS, and Linux on Node.js 20 and 22. Releases are manually authorized, checksum-protected, and immutable.

Stop the Manager safely:

node scripts/taskmaster.mjs manager stop --json

Start with skills/eric-task-master/SKILL.md. Technical details live in ARCHITECTURE.md, docs/MCP.md, and docs/RELEASE-GATE.md.

License

MIT. Use, modify, distribute, and build open or proprietary Skills and Task Packs on top of Task Master.