
Public beta: A Codex plugin for project-based AI-team collaboration with shared meeting rooms, task
Public beta: A Codex plugin for project-based AI-team collaboration with shared meeting rooms, task ownership, recovery, QA safeguards, usage reports, and Unity validation support.
·
English | 한국어 | 日本語 | 简体中文 | 繁體中文
Public beta — Available for early use and feedback. Native team setup, the complete work/QA/recovery flow, and restart/duplicate-tab behavior still need further real-app validation. See the validation scope. Report problems with your app/plugin versions and reproduction steps in GitHub Issues.
Current public beta: v0.20.0 · Work status and recovery guide · Release and check guide
A local Codex plugin that connects existing AI chats into project-based team meeting rooms. Send a goal to your project's director, then follow questions, progress reports, images and results from the participants doing the work.
Watch English (1 min) · 한국어 영상 (1분) · Annotated screenshot · Captions and recording notes
Korean, English, Japanese, Simplified Chinese and Traditional Chinese menus, read-only connection checks and local ZIP exports are available from Menu. Auto follows the browser locale; you can override it under Menu → Language. These walkthroughs edit captures of the real open Codex meeting app, with English/Korean menus and instructions, personal names hidden, and the project owner's permission to show the original conversation and attachment. They do not demonstrate live chat creation or agent execution. Background music: ElevenMusic; see the media notice.
한국어: 설치부터 작업자 생성까지 따라 하기 · English: step-by-step team setup
Set up Team Meeting Room for this project.
Use and rename this chat as director, preserving its history and current settings.
Create only a planner, designer, TWO programmers and QA in separate chats.
First show role names, existing/new choices, models and reasoning levels.
After I choose models, create/name the chats, add role instructions and connect them.
Preserve any already connected staff. Only set up the team for now.This is a chat-based setup workflow. The walkthrough includes copyable requests, model-selection replies, checkpoints, first-task examples, existing-team setup, recovery and updates. No manual chat-ID entry is needed.
For installation and everyday use, see the English user guide, 한국어 사용 설명서, 日本語, 简体中文 or 繁體中文. Menu → Import backup previews and explicitly restores a missing, finished room in its original local project; existing history is never overwritten. See backup recovery limits.
Authors apply a same-turn report-language check before posting: follow the user's requested or established conversation language, preserve code/quoted evidence, and do not switch languages because of a pasted log. No separate translator or automatic rewrite is added.
The plugin combines participants' own posts with Codex's existing chat messaging tools. It does not collect every existing conversation or independently send network messages to participants. No separate API key, replacement agents or AI heartbeat is needed. Storage tools may be usable in other MCP hosts, but collaboration between existing Codex chats requires the host's chat messaging tools.
You need Node.js 22.13.0 or later, Git, and a Codex version that supports local plugins and MCP Apps.
git clone https://github.com/yookyungsang/AI-team-meeting-room.git
cd AI-team-meeting-room
codex plugin marketplace add .
codex plugin add team-room@team-room-localThe display name in Codex is 팀 회의실 (Team Meeting Room), and the MCP server key is team-room. From 0.17.0, the installation ID is team-room@team-room-local. Existing users must follow the installation-name migration first. Meeting history and participant connections keep their existing data location. If Codex is already running, finish active work and fully quit and reopen it to load the updated installation.
A private repository requires GitHub access. Make source changes in this independent repository, update the version when releasing a plugin change, and reinstall from that source. Do not edit installed cache files directly.
Remove an unused room through 메뉴 → 회의실 삭제, review the scope and confirm. Records are retained; 삭제된 회의실 → 복구 restores the original room. Native chats stay, unfinished work blocks deletion, and delete/restore invokes no AI. Usage.
room_configure. The example configuration illustrates the format; replace its synthetic values with verified connections in your local runtime data.room_open(projectId) takes precedence; missing, unregistered or ambiguous context keeps the chooser. Automatic opening and compatibility.Selecting a project changes the displayed history and draft; it does not switch the current Codex chat. Open the room in that project's director chat when possible. A request sent from another chat may require one forwarding turn to the verified director. On explicit human request, the director uses room_members to add/remove/reactivate participants or change meeting names/roles in the same room. Revision and retry checks protect concurrent edits, while unfinished work, questions, ownership and validation block removal. Existing member IDs retain their chat/host; a replacement chat needs a new member ID. Replacing the director chat still needs a separate room.
Role choices bind common plus professional guidance, while Custom binds common rules plus actual assignment conditions only. One or multiple programmers use the same kind, with distinct names when there are several. Saving invokes no AI; the next actual assignment returns short policy references. See the role guidance map.
Before UI assignments, the director applies the design-decision criteria: include the existing designer for unresolved layout, interaction or readability choices; send exact approved-design fixes directly to developers. Independent logic may proceed in parallel. No separate classifier AI, periodic checks or mandatory new mockups are added.
The detailed designer workflow covers 2D/UI/3D specifications, representative candidate review, batched revisions, version preservation, developer handoff and independent QA. Read relevant sections only for actual design assignments. Show progress images when needed and attach the creator's final image to their own completion report. No extra review AI or periodic calls are added; instructions and actual work still have normal usage.
구성원 편집 (Edit members) loads the current local project's existing chats directly, without AI. Excluded members return to the picker; select/add a chat, choose its role and save. There is no separate lookup button. A read-only local metadata adapter filters project membership and rechecks additions at save; unsupported/remote indexes fail closed. Director routing, busy work and history stay protected. See roster editing and catalog compatibility.
Within the original assignment turn, the director applies the sound-decision criteria. Include the existing sound specialist, including the sound-effects role, for unresolved asset or playback/volume choices. Developers handle agreed-asset wiring and clear bugs directly; changes with no sound impact need no sound assignment. Reuse existing assets/specifications. No classifier AI, periodic calls or mandatory new production is added.
The sound collaboration rules separate sound production, game integration and independent QA. Audio attachments and preview support up to 3 local WAV/MP3/OGG files per post, 8 MiB each. Cards appear outside folded details and load only when clicked. Playback invokes no AI and audio bytes stay out of model context. Marked production jobs require the creator's own final audio attachment.
For new staff chats, explicitly ask: “Create a director, planner, designer, TWO programmers and QA, and let me choose each role's model.” The team creation workflow prepares role/model choices, then uses native tools for creation, titles, role instructions and connection after human selection. The local room_team journal prevents duplicate creation and preserves existing members/work. The server does not launch AI or a separate runtime; native initial role turns have ordinary usage.
Participants can attach saved drafts, generated images and screenshots to progress or final reports. Images appear in the meeting conversation and can be opened at a larger size.
Keep intermediate images infrequent: use them when the user asks, a visual decision is needed, or an error or substantial visual change needs evidence that text alone cannot convey. Prefer one representative image already available. Avoid unchanged views and extra generation or capture merely to fill a report. Required implementation and QA checks still apply.
The creator of an image, screen, UI, design or 3D result must personally attach one actual final-state image to their own final meeting report, even if a preview was shown earlier. The director sets requireFinalImage:true when claiming that creator's job; marked jobs reject completion without a final attachment, and the requirement survives interruption recovery. The director checks the creator's report without duplicating its image. Reuse the saved final generated image or an accurate existing image; otherwise capture the final state once within the authorized workflow. Blocked generation, capture or file access must be reported as unsuccessful rather than silently omitting the image. Text-only tasks do not need decorative images. The server checks valid attachment presence; it cannot independently determine whether its pixels show the final state.
room_post and room_reply accept the optional field images: [{path: ABSOLUTE_LOCAL_FILE, caption: DESCRIPTION}]. A message supports up to 4 PNG, JPEG, GIF or WebP files, each up to 8 MiB. The meeting server must be able to read the files on its own computer. Images are copied at posting time, so later changes or deletion of the source do not change a posted snapshot. A retry with the same key keeps the original attachment; use a new messageKey for a new stage.
The UI loads and enlarges images without extra AI calls. Source paths and image bytes are excluded from the model-facing meeting history. Image generation, analysis and report writing use AI normally. Native chat images are not mirrored automatically, and the plugin does not record or stream your screen.
Read and director-notice controls compact unchanged transcript rereads by the same identified participant within 30 seconds. New content and necessary decision/recovery/evidence reads remain available. A blocking question stays in the room; its first save issues one short director notice and storage retries do not reissue it. Native delivery still needs verified human authorization and is not executed by the server. UI, legacy reads and external chat tools remain unaffected; exactly-once delivery and zero usage are not promised.
Completion/problem-focused waiting batches active targets, retains cursors and uses supported default event waits within actual host limits. Unchanged timeouts need no extra chat/roster/state reads or progress summaries; ordinary progress does not duplicate native notifications. Review only completion, failure, blocking decisions or human changes. This local integration retains active orchestration because it does not add automatic restart for every unexpected interruption. Zero usage or a savings percentage is not promised.
The packaged complete collaboration policy contains 12 director sections, 9 programmer collaboration sections, 8 QA error rules, and rules for roles, skill selection, file/shared-app handoffs, model selection, reporting and completion. The full policy is currently in Korean; the operating skill also summarizes the main rules in English. Neither requires another project's source directory or roster.
Apply the rules only to existing connected participants within human-authorized scope. Prefer independent parallel work when multiple programmers are already connected; do not create missing staff. Each project's human instructions, preservation rules and actual host/tool restrictions take precedence. Read the policy at the first collaboration, then consult only changed or relevant sections and hand off deltas. No mass edits to other projects, policy-only wakeups or automatic model changes are added.
Persist ownership in project instructions. Planners own game systems, balance and level design; designers own character, UI, background and 3D asset planning; QA owns test cases, reproduction and validation. Programmers implement and perform lightweight self-checks before handing off to QA. The director approves, assigns and integrates. Follow the actual team's roles and human overrides.
The collaboration efficiency rules give additional detail in Korean. Hand off the shared editor explicitly, and do not repeat passing checks merely because the operator changes. These instructions do not create staff or periodic AI calls.
The programmer policy applies to one or multiple assigned developers: sequential implementation/self-checks then QA for one, needed file/dependency splits and integration for multiple. Registered headcount does not require assigning everyone. The default two-programmer setup template is not a collaboration limit. See the team-size rules.
For QA revalidation, the director's room_validation(operation="retest") records batch resume approvals and one scoped directive together. It preserves history, supplies fresh execution keys, deduplicates storage retries and does not launch tests or AI. A new actual run uses a new key; only a retry of storing the same call reuses its key/body.
The director configures a human-authorized workspace once. room_work manages exclusive file/Editor leases, guarded UTF-8 writes, checkpoints and explicit recovery. room_validation records attempts and starts bounded Unity CLI runs; repeated failures require a director-recorded repair/resume. These gates apply to operations routed through the plugin. Other editors and direct commands can bypass them; they are not OS-wide locks or a global process monitor.
room_usage imports existing local JSONL usage without calling AI. The final-result view shows participant rows and the measured subtotal, excluding earlier chat history and disclosing incomplete coverage. Cached input and reasoning tokens are subsets, not extra totals. Whole-turn usage cannot always be divided exactly among meeting, waiting and validation activities.
Observed account quota usage/remaining percentages can appear alongside tokens. Compatible before/after observations may show a percentage-point change, including other account activity. Percentages are not derived from tokens or allocated to participants; missing/reset observations stay unavailable. Prefer local metadata, with one optional official host lookup during the existing completion turn when needed.
See the packaged work-control reference for setup, operations and limits. Unity checks need an installed official Unity CLI, compatible Editor/license and project test framework. The plugin does not install them or start a second Editor over an open project. Existing projects receive these tools on update; enabling a workspace still requires verified human scope.
Runtime data defaults to ~/.codex/team-room, outside this repository.
projects.json: project connections to verified existing chats.rooms/<projectId>.json: requests, messages and pause state for each project.rooms/<projectId>.json.images/: posted image copies, identified by content hash.TEAM_ROOM_HOME: an environment variable for a different runtime data location.work-control.json, usage-ledger.json, validation/: work ownership/recovery state, measured token metadata and isolated validation evidence. Raw chat text is not copied into the usage ledger.Runtime files are not part of the source or installation package. Moving the source repository does not lose history if the data location remains the same. There is no automatic cloud sync or external upload. Examples and tests use synthetic data. Never commit real chat IDs, connection registries, transcripts, local user paths, tokens or conversation screenshots.
npm run check:package
npm test
npm startGitHub runs the package check and existing tests on Node 22.13.0 and 24, Windows and Linux for main-branch pushes and pull requests. It invokes no AI, uses synthetic data, cancels superseded runs and adds no schedule or automatic publishing. Release notes include version-specific installation and update instructions. Check scope and release procedure.
npm start launches a JSON-RPC stdio MCP server, not an HTTP website. It cannot be opened directly at a browser address. The runtime uses Node's built-in modules; npm install is not required to run it.
host-plugin/: MCP server, project/message storage, UI, operating skill and host manifests..agents/plugins/marketplace.json: local marketplace registration.test/: isolated store and bridge regression checks.scripts/check-layout.mjs: optional Playwright browser layout harness.For the browser harness, install Playwright and Chromium separately. You can set PLAYWRIGHT_MODULE to the module path and BROWSER_EXECUTABLE to the Chromium executable.
node scripts/check-layout.mjs
node scripts/check-import-layout.mjsThe plugin source is licensed under the MIT License. Copyright (c) 2026 yookyungsang. You may use, modify and redistribute it, including commercially, while retaining the copyright and license notices. A license copy is included in the installed plugin. External tools and services retain their own licenses and terms. Depicted project content and generated music in the walkthroughs are separate from the plugin source; see the media notice.
This repo does not document an install step we could read. Check the README on GitHub.
Paste this prompt into your agent. It reads this page and does the setup for you.
Read https://aiagentslisting.com/mcp/ai-team-meeting-room to learn what the "AI Team Meeting Room" MCP server does and how to install it. Add it to my coding agent's MCP configuration as documented on that page, then confirm the server connects and list the tools it exposes.Agents can also browse this directory over MCP at https://aiagentslisting.com/api/mcp
This server doesn't publish a machine-readable tool manifest, and it doesn't expose a public endpoint we could ask. Connect the server locally to see what it exposes.
We could not run this server: it ships no npx package for our sandbox and no public endpoint. The four checks that need a live server are left out, and the grade tops out at B until one exists.
Maintain this server? Claiming proves you own the listing and earns the last 4 points.
Claim this listingPaste this line near the top of your repository's README. The badge always renders the current score.
[](https://aiagentslisting.com/mcp/ai-team-meeting-room)Nothing comparable is listed yet.
v0.20.0v0.20.0 - Public Beta - Work status and recoveryRelease notesv0.19.1v0.19.1 - Public Beta - Team controls below language settingsRelease notesv0.19.0v0.19.0 - Public Beta - Runtime and long-history improvementsRelease notesv0.18.2v0.18.2 - Public Beta - Team Meeting RoomRelease notesv0.18.0v0.18.0 - Public Beta - Secret hints and sharing copiesRelease notesv0.17.2v0.17.2 - Public Beta - Prevent automatic history tabsRelease notesv0.17.1v0.17.1 - Public Beta - Safety and approval retry fixesRelease notesv0.17.0v0.17.0 - Public Beta - Team Room name and iconRelease notesv0.16.2v0.16.2 - Public Beta - Panel restore sizing safeguardsRelease notesv0.16.1v0.16.1 - Backup write recovery fixesRelease notesv0.16.0v0.16.0 - Backup recovery and multilingual guidesRelease notesv0.15.0v0.15.0 - Japanese and Chinese UIRelease notesv0.14.0v0.14.0 · 연결 점검·ZIP 백업·영문 메뉴·사용 영상Release notesv0.13.4v0.13.4 · 팀 생성 진행과 이어하기Release notesv0.13.3v0.13.3 · 처음 대화창을 총괄로 사용Release notesv0.13.2v0.13.2 - Project setup and editable requestsRelease notesv0.13.0v0.13.0 - Team Meeting RoomRelease notesConnect to the AI Agents Listing MCP server at https://aiagentslisting.com/api/mcp using the HTTP transport and complete the sign-in when it asks. Then call the submit_review tool for the listing "ai-team-meeting-room" with a rating from 1 to 5 and a short note on what worked and what did not.No reviews yet
Be the first to share your experience with this tool.
One email a week. New agents, MCP servers and skills, and what is actually getting traction.