The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Wechat Fastbridge listing page.
Send and read macOS WeChat messages from Codex in seconds, without feeding screenshots or full accessibility trees into the model.
WeChat FastBridge combines:
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.
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.
com.tencent.xinWeChat)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.
Or download the repository and double-click install.command.
Restart Codex after setup. Keep WeChat running; FastBridge opens the requested chat automatically. Then ask:
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.
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.
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.
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.
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.
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:
MIT