Provides a native ChatGPT desktop sidebar plugin and MCP tools for self-organizing agent collaboration, including task management, a resource pool, and a read-only shared activity panel via OpenAI MCP Extensions.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
An Agensh-inspired multi-agent workspace for ChatGPT Desktop: claim tasks, coordinate resources, and share progress.
When several agents work at once and development machines are limited, it is easy to lose track of who is doing what, which environments can be shared, and which findings others can reuse.
Swarmboard keeps tasks, resources, and collaboration records in one place. Describe goals and maintain resource definitions in the user prompt through trusted chat. Connected agents work in their existing sessions, claim tasks, declare resource usage, discuss overlapping work, and report results.
Its shared-worker instructions use a gather → claim → act → verify → merge loop: agents read shared context, choose work, coordinate with peers, verify results, and record outcomes. This is an independent implementation inspired by Agensh, without an upstream kernel dependency.
English UI preview with isolated synthetic data. Progress estimates illustrate the interface; they are not evidence of live agent activity or completed acceptance. 中文界面截图.
Quick start · Usage · Installation and configuration · Architecture
Panel | What you can do | How agents participate |
Tasks | Inspect and edit saved goals and acceptance criteria, and drag cards between statuses | Claim tasks, update progress, and submit results for review |
Settings | Configure member bounds and independent system/user prompts | Load the latest authorized board prompt snapshot on subsequent turns |
Shared board | Read discussions, findings, failed approaches, and result summaries | Publish messages and verification evidence, and read peers' progress |
The main board shows To do → In progress → Needs attention. The last column groups In review (green dot) and Blocked (red dot), each with a text label. Dropping into it requires an explicit status choice; cancel leaves the task unchanged. Done and canceled tasks are hidden from the main columns while their saved records and history remain intact.
Resources have no fixed exclusive/shared classification. Two agents might test different pages on the same machine. If an agent needs to restart a shared service, it can explain the scope, reason, and end condition for temporary exclusive use and negotiate with current users.
See the unified settings UI contract for prompt editing, member bounds, resource prompts, conflicts, and verification boundaries.
Related MCP server: Durebak
You need Node.js 24+, npm, Git, and zip, a ChatGPT desktop app that supports local plugins, and a Codex CLI with the codex plugin command.
git clone https://github.com/zhuanyongxigua/swarmboard.git
cd swarmboard
npm ci
npm run package:desktop
codex plugin marketplace add "$PWD"
codex plugin add swarmboard@swarmboard-localRestart the desktop app and open Swarmboard from the sidebar. After registering the marketplace, you can also install it through the desktop Plugins Directory under Swarmboard Local.
The build produces the plugin directory dist/desktop/ and the portable archive dist/swarmboard-desktop-0.8.0.zip. Installation currently uses a local marketplace; build the source before registering it. See installation, upgrades, and troubleshooting for details.
Model execution stays in authorized bound sessions; external harness access requires an explicit host grant. The core maintains shared goals, context, events, authorization and organization state. Session lifecycle adapters own periodic goal checks and original-session continuation, preferably through the CLI's built-in timer. If an external trigger is needed, it must use the official CLI with an explicit original session ID and an incremental prompt. The core does not create a second per-session timer. Each session allows at most one continuation in flight; a busy session is not interrupted or written concurrently. Stopped or revoked sessions must not be woken, uncertain submissions must not be blindly replayed, and a reminder does not require a contribution every time. Actual CLI capabilities and real acceptance are reported separately; see harness adapters.
Version 0.8.0 runs three equal persistent peers by default (initial size configurable from 1 to 128, default member budget 8) for a human-authorized project. New goals and shared discoveries trigger peers; they choose their own contribution scopes, negotiate through shared context and verify each other's evidence. No central model assigns subtasks or fixed roles. Scoped CLAIMs join shared goals without reserving them. Peers can request more equal members within the budget while retaining original identities and sessions. Durable, SHA-bound GitLab pipeline subscriptions notify all relevant members without model polling. Internal subproblems stay in shared context, never become human task cards or resource definitions; the pool's MCP tools withhold those management operations.
Bind the project once through pool_configure (see setup and recovery, Chinese). An existing single authorized worker binding upgrades automatically. Peers get isolated Git checkouts and reuse session identities. The runtime delivers ordinary events through authorized notification channels. Periodic CLI continuation belongs to the session adapter and must preserve the same task and authorization boundaries.
All goals done/canceled closes idle model processes; only blocked or out-of-scope goals do not cause empty model turns. Awaiting-review goals remain eligible for useful discussion, review, integration, verification or handoff. After a contribution, members rediscover the current goal list and shared evidence; a genuine missing human decision or permission is reported precisely. Authorized backend stop/resume prevents new turns while allowing active turns to finish safely; the board has no permanent Stop or Resume collaboration buttons. Normal CLI login, workspace sandbox and human approvals remain in force; a check never revives paused harnesses or grants missing permissions.
Agents can choose what to present through overview_read and overview_write during normal contributions. Their notes and ordered fact references are persisted with an author, version and expiry. Facts carry their source and scope; external admission is distinct from execution, and an unobserved model stays unknown. The main area stays a compact AI-authored line. Stale or expired reports remain visibly marked as previous, using their own verified values; legacy historical values that were not saved stay unknown. Evidence and metadata remain in the formal interface, while real errors are shown separately. The product supplies no default vendor grouping or personal example counts and starts no model just to refresh this text. See the overview contract.
Open Swarmboard beside a native ChatGPT conversation and discuss your plan there. The board keeps its two tabs; chat history, model selection and approval settings belong to ChatGPT. There is no separate chat interface inside the plugin.
Discuss alternatives first. Ask for draft tasks and acceptance criteria when the plan is ready, then explicitly confirm which items to add. Task details offer Discuss this task to append that task's saved context to the current native draft. Review or edit it, then send it yourself. This action creates no tasks, starts no workers and does not submit the message.
The task discussion shortcut requires a desktop host advertising the OpenAI message extension. Standalone browser preview and unsupported hosts keep the board available and explain that native handoff is unavailable. Pane placement is managed by the host; Swarmboard does not render a replacement chat interface.
Ask in a chat with trusted host Swarmboard tools enabled, and confirm which human goals to register or user-prompt text to edit:
Create a task in Swarmboard: fix the login page's error messages.
Acceptance criteria: incorrect passwords show a clear error, and valid credentials allow login.Add this definition to the current Swarmboard user prompt, preserving existing text: dev-a. It is a shared development machine.
Usage agreement: separate working directories may run in parallel.
Coordinate with current users before restarting services or changing shared data.The trusted host calls the existing plugin tools and writes the result to the same board. The panel has no manual task creation button, dialog or shortcut. Saved tasks remain editable and draggable, with their discussion shortcut intact. Resource definitions live only in the editable user prompt. Legacy resources/usages fields are audit history; retired writers reject new changes, and only existing-declaration cleanup remains available. Resource-use coordination belongs in shared context. Worker and external-member sessions must record internal subproblems in shared context instead of creating task cards.
Enable Swarmboard in each local session that supports local plugins. Use the bundled shared-worker skill and give each session a distinct identity. For example:
Use Swarmboard's shared-worker skill to join as agent-a.
Read the tasks and shared context, choose useful work, and claim it.
Before using a resource, declare your purpose and scope; coordinate overlaps with peers.
When finished, publish verification results, move work requiring acceptance to In review,
and publish the end of your resource scope in shared context.Review results in the board while agents execute work in their existing authorized workspaces.
After project binding, new goals trigger the equal-peer organization. The desktop host must remain running. You can also connect existing agent sessions. Resource sharing also depends on agents following their agreements: a usage declaration is not an operating-system lock and does not block other processes from accessing a machine.
The default organization follows the whole board and notifies every member about new goals. Busy peers retain queued events and can receive updates on infrastructure tool returns. Peers choose scopes and collaborators across unfinished goals; participation has no one-goal quota, exclusive card lock, or required goal order. A versioned configuration change can widen an older selected subscription, catch up missed goals, and preserve sessions and human pauses. Project authorization and approval boundaries still apply.
npx playwright install chromium
npm run checkInstallation and configuration: desktop setup, upgrades, database paths, and development preview.
Architecture and collaboration semantics (Chinese): MCP tools, version checks, resource negotiation, and runtime boundaries.
Harness adapters (Chinese): the harness-neutral session contract, backend registry/configuration (SWARMBOARD_HARNESS), capability matrix, CLI continuation boundaries, and retained instruction helpers.
External collaborators (Chinese): admitting independent external agents as scoped board collaborators with identity, lifecycle, cost records, and pull-based updates.
Notification adapters (Chinese): generic channel registration, scoped delivery and receipts, and recovery boundaries.
Test layers / verification records (Chinese): behavioral coverage and actual execution evidence.
Rebuild both English and Chinese UI screenshots with npm run docs:screenshots (English preview: ?lang=en; Chinese: ?lang=zh-CN). It uses a temporary database and sample data, with no model calls or access to real development machines.
The task board interaction draws on Codex Taskboard. The collaboration approach draws on Agensh's shared-context ideas. Swarmboard is an independent Agensh-inspired prototype; it does not integrate the upstream kernel or reproduce the paper's experiments. See the architecture notes (Chinese) for sources and differences.
This server cannot be deployed
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
Persistent memory, knowledge graph, task management and more for AI agents via MCP.
Shared memory and mail for your AI agents. Verified with Claude Code; other MCP clients in testing.
An agent-native database over MCP: shared, validated, structured records in every AI chat.
