Douyin Controlled MCP

An experimental, GPT-first (not GPT-only) MCP adapter for reading and carefully operating Douyin and Creator Center through structured IDs, DOM state, and verified browser state. Visual or Computer Use operation is a fallback when a structured route cannot safely identify the target.

This repository is an early v0.1.0-alpha candidate for a small tester group. Douyin UI and anti-abuse behavior can change without notice. Readback, confirmation gates, and stable IDs reduce risk but do not make automation risk-free.

First run on Windows

Prerequisites: Windows 10/11, Node.js 20 or newer, Chrome or Edge, and a Douyin account that you can sign into yourself.

1. Run START_HERE_SETUP.cmd.
2. Run START_BRIDGE.cmd and connect the displayed MCP URL to your AI client.
3. Sign in to Douyin in the dedicated browser and open your own profile or Creator Center.
4. Tell the AI: “请帮我完成抖音首次配置。”
5. The AI checks configuration, detects the current account, asks you for a local alias and write permissions, creates ignored private configuration, and validates it.

You should not paste cookies, tokens, uid, or sec_uid into chat. See AI_SETUP.md for the exact machine-readable flow.

Connection modes

START_BRIDGE.cmd uses a Cloudflare Quick Tunnel by default. It is the easiest Alpha path, but its public URL can change after a restart.

If a tester wants a stable endpoint, a local AI may configure that tester's own Tailscale Funnel after the bridge is running. Tell the AI: “Configure an optional stable Tailscale Funnel endpoint for this MCP.” The AI must follow AI_SETUP.md, discover the installed Tailscale CLI and its current syntax, verify the endpoint, and keep the generated hostname and access token only in ignored runtime/ files. A hostname, Tailnet name, or access token from another installation must never be copied from documentation or committed to Git.

Stable public access is optional. Local-only access and the default temporary tunnel remain supported. If Tailscale requires the owner to sign in, enable Funnel, or approve a policy change, the AI must pause for the owner instead of bypassing that authorization.

Safety boundary

• Structured tools are preferred; screenshots and visual operation are fallback paths.
• Initial write permissions default to off. The AI must ask before enabling comments, replies, or publishing.
• Sending, publishing, deletion, account changes, payments, identity checks, and other sensitive actions require the owner to confirm or take over.
• Do not connect multiple AI clients to the same browser session for write work.
• Runtime configuration, browser profiles, tokens, cookies, logs, screenshots, SQLite state, and generated connector URLs are excluded from source and release packages.
• The old third-party EchoLens transcription route is retired. Optional transcription uses local faster-whisper only.

Development verification

npm ci
npm test

npm test runs TypeScript checking, build, deterministic fixtures, recovery/idempotency fixtures, HTTP security fixtures, and first-run configuration fixtures. Live Douyin behavior still requires owner-authorized manual acceptance because it depends on a logged-in account and the current site UI.

Release and provenance

The application code is licensed under MIT; see LICENSE. Dependency and optional-tool notices are in THIRD_PARTY_NOTICES.md. The owner-provided origin statement is recorded in PROJECT_ORIGIN.md.

Public release must use a fresh sanitized snapshot, not this private repository's historical Git objects. See SECURITY.md before packaging.

Known limitations and deferred architecture are tracked in KNOWN_ISSUES.md and ROADMAP.md.

Related projects

For the unified entry point to the Douyin, Windows PC, and Android phone MCP projects, see mcp-tools-link-hub.