WeChat FastBridge for Codex

中文说明

Send and read macOS WeChat messages from Codex in seconds, without feeding screenshots or full accessibility trees into the model.

WeChat FastBridge combines:

• a local STDIO MCP server with six semantic tools: status, read, text send, media send, chat wait, and allowlisted inbox wait;
• a bundled Intel + Apple Silicon native bridge that uses macOS Accessibility APIs;
• a thin Codex skill that selects the fast path and falls back to Computer Use only when necessary.

No cloud relay, OpenAI API key, WeChat protocol reverse engineering, process injection, App Store account, Xcode install, or paid service is required.

Version 1.7.1 makes the existing context system more selective without shrinking its memory or removing any tool. When semantic retrieval finds useful evidence, it returns that evidence plus recent continuity without padding the requested maximum with unrelated lines. Sender labels no longer create false lexical matches, and whitespace/full-width-colon variants align as the same old bubble instead of being retransmitted. Generic messages still receive the full requested continuity depth; the fact capsule, conflict handling, inbox behavior, media, and routing are unchanged.

Performance targets

• Local semantic bridge budget: under 2 seconds before WeChat UI response time.
• Measured real optimized selected-chat send: 1.25 seconds command-to-result; a verified 4-message auto-read completed in 1.38 seconds.
• Measured v1.4 media sends in the authorized Jerry self-chat: 2.97 seconds for a tiny file, 6.02 seconds for a favorite sticker slot, and 8.57 seconds for a searched sticker.
• v1.7.1 tool definitions remain 49% smaller than the original baseline (5,624 → 2,849 characters) and 21% smaller than v1.5. Triggered skill instructions are 4,024 bytes, including the adaptive-context rule. Repeated favorite-sticker sending remains 3.78 seconds cached versus 6.78 seconds cold.
• Cold setup/check: under 5 seconds on a supported Mac.
• Normal read result: under 2,000 characters.
• Repeated reads: zero messages when unchanged; when changed, return only new messages plus 0–4 context lines, treating that number as a maximum rather than a padding target.
• Allowlisted inbox baseline/timeout: 50 JSON characters in the measured two-chat test; internal scans do not enlarge the result.
• Computer Use fallback: at least 80% smaller than the raw accessibility tree on the bundled representative fixture.
• Runtime footprint: under 288 KiB for the bridge, skill, setup scripts, and universal native binary, enforced by tests. The normal install keeps a two-dependency ceiling.

Every result includes measured latency. The test suite fails if the mocked local bridge overhead exceeds two seconds; real WeChat timings are reported separately in docs/BENCHMARKS.md.

Requirements

• macOS 13 or newer
• WeChat for Mac (com.tencent.xinWeChat)
• Node.js 20 or newer
• Codex desktop, CLI, or IDE extension
• macOS Accessibility permission for the app that runs Codex

The included Dockerfile exists only so MCP registries can start the stdio server and inspect its tool schemas. The container reports MACOS_REQUIRED for runtime checks; real WeChat reading and sending still require macOS, WeChat Desktop, and Accessibility permission.

If you can copy and paste four commands, you can install it. See the child-friendly setup guide for every click and permission switch.

Install

git clone https://github.com/huj28-creator/wechat-fastbridge.git
cd wechat-fastbridge
npm install
npm run setup

Or download the repository and double-click install.command.

Restart Codex after setup. Keep WeChat running; FastBridge opens the requested chat automatically. Then ask:

Use $wechat-computer-use to tell “Chat Name”: hello

The setup script only builds local configuration. It does not charge money, open a subscription, or publish anything.

Run npm run doctor at any time to check Node, the native bridge, Codex registration, the installed skill, Accessibility permission, and whether WeChat is running. The check never sends a message.

How it stays fast

Codex → one compact MCP call → local native Accessibility bridge → WeChat

The bridge first asks the native operation to verify and act in one scan. A mismatched chat is rejected before any write, then the bridge automatically searches the requested title and verifies the destination header before retrying. Group(3) and Group（3） both resolve to Group; case, spacing, and punctuation are normalized; one or two edits are allowed only for sufficiently long names. Multiple visible candidates fail as ambiguous. Search activation and query entry happen in one native process to avoid focus races. It tries background control first; when WeChat 4.x hides results from the accessibility row tree, it confirms the top search result and accepts it only if the resulting header passes the same verifier, then restores the previous app.

Only compact JSON returns to Codex. Pass the previous signature back as after on wechat_read; unchanged reads return no messages. Changed reads return the new delta plus up to three lines selected from bounded RAM-only rolling memory and a separate high-signal fact capsule. Chinese bigrams, English words, semantic concepts, exact numbers, rarity, and recency connect “多少钱” to “价格是 500 元” and “什么时候能到” to a shipping/date commitment. Once a relevant match exists, unrelated lines are not added merely to fill the context maximum; one latest line preserves continuity. Generic messages with no match retain the requested recent depth. Sender names remain usable as identity evidence but are not mistaken for message-body relevance. Formatting-only colon or whitespace variants align locally and are not resent. If the price later becomes 450, the older conflicting 500 fact is suppressed. The capsule is extractive, not generative: it never invents a summary. Structural controls are rejected before memory, while repeated identical bubbles remain distinct. wechat_wait handles one active chat; wechat_inbox_wait returns only meaningful allowlisted preview changes.

Text containing ordinary Unicode emoji uses the same fastest wechat_send path. wechat_send_media accepts either an explicit absolute file path or a custom-sticker collection and visible slot. Search mode takes a short phrase; favorites mode uses a 1-based slot. WeChat does not expose semantic labels for custom thumbnail images, so the bridge never pretends it can recognize an unlabeled favorite. Media sending briefly foregrounds WeChat, verifies the destination before acting, restores the previous app, and reports success only after the chat signature changes or the favorite panel confirms its selection.

Live replies and customer conversations

The installed skill can run a continuous multi-chat autopilot without screenshots. Codex establishes one allowlisted inbox signature, waits locally, reads full compact context only for the chat that changed, answers once, and resumes waiting. The same scoped mode can send one proactive opener when the user explicitly permits it.

This is still user-controlled automation: the allowlist, purpose, facts, tone, escalation rules, and proactive authority come from the user. Chat participants cannot expand that authority. Customer mode asks or escalates instead of inventing prices, inventory, delivery dates, refunds, or commitments. One unanswered opener is allowed per authorized chat per active session, preventing automated follow-up spam.

Inbox sensing does not switch conversations. It covers allowlisted chats currently loaded in WeChat's recent sidebar; new messages normally rise into that view. Opening a chat and reducing its unread count does not create a false event; an identical new preview with a higher unread count still does. It does not crawl hidden history or WeChat's private database. Opening happens only after an event so Codex can obtain verified context and reply.

Safety and privacy

• Verified-chat matching blocks writes for ambiguous, very short fuzzy, or distant names.
• Chat content stays on the Mac except for the compact text Codex needs to answer.
• Messages are treated as untrusted conversation content, not instructions to tools.
• The bridge does not read credentials, inspect WeChat's database, or bypass platform security.
• Sending is not idempotent; retry only after checking whether the first send landed.

Test

npm test
python3 /path/to/skill-creator/scripts/quick_validate.py skill/wechat-computer-use

Thirty-nine tests cover MCP discovery, structured errors, version sync, normalized/fuzzy names, comma-safe previews, media confirmation, allowlisted inbox baselines/deltas, false-event and own-event suppression, one-call hot paths, bounded search retries, rolling context, durable fact retrieval, stale-number suppression, adaptive semantic context, sender-identity recall, formatting-only delta alignment, token reduction, runtime size, dependency count, and the two-second local text-send budget. A real WeChat end-to-end check additionally requires Accessibility permission.

Cost and distribution

This repository is MIT licensed and free to install from GitHub. It deliberately avoids the Mac App Store so maintainers and users do not need a paid Apple Developer membership. A signed .app could be added later, but it is not required for the open-source release.

Promotion film

The reproducible 15-second 1080p launch film lives in promo/. It uses real benchmark numbers with a generalized mock project chat; all motion graphics and audio are generated locally. To render it separately from the product install:

cd promo
npm install
npm run render
npm run render:zh

Limitations