README
¶
AgentHub — Multi-Agent Collaboration Platform
AI Full-Stack Challenge entry · Team 启灵 (Qiling) — 齐紫瀚, 熊子恒, 张桐铖
中文版:README_CN.md
AgentHub makes "driving a team of AI agents" feel as natural as a group chat: agents are room members; @-mention the lead agent to auto-decompose and dispatch work, or @ a specific member to assign it directly. Artifacts land in a shared room workspace, and the whole collaboration is visible in the chat stream in real time. The platform plugs in multiple agents through one adapter layer — memoh (built-in framework agent in this repo), Claude Code, and Codex (CLI-backed) — coordinated by an Orchestrator DAG engine.
This repository is built on the memoh full-stack platform (Go backend + Vue web + Electron desktop + containerized agent workspaces). AgentHub is its multi-agent collaboration capability; memoh is one of the agents that run inside it.
📦 Deliverables (3 documents)
The three final documents live in deliverables/ as PDFs:
| Document | File | Contents |
|---|---|---|
| Product Design | agenthub-product-design.pdf |
Positioning, the group-chat mental model, multi-agent failure modes & mitigations, target users & scenarios, core UX decisions, feature matrix, roadmap |
| Technical Design | agenthub-technical-design.pdf |
Layered architecture, data model (full DDL), orchestration state machine & scheduling, two-tier Planner, unified adapter layer, credential chain, event projection, three core sequence flows, HTTP API, test gates |
| AI Collaboration Record | agenthub-ai-collaboration.pdf |
Human–AI turn-based model, Rules / Spec / Plan / Skill conventions, three deep collaboration cases, real session archive (19 sessions / 277 user turns / 2582 tool calls) |
Rubric navigation
| Dimension | Weight | Where to verify |
|---|---|---|
| AI collaboration | 30% | AI Collaboration Record + git log (R1–R9 feature commits) |
| Feature completeness | 25% | Run the demo below; feature matrix in the Product Design |
| Output quality | 20% | Demo video + screenshots below |
| Code understanding | 15% | Technical Design (architecture / state machines / sequences) + internal/agenthub/** |
| Innovation & product sense | 10% | Product doc: failure-mode mitigations, pinned long-term context, shared workspace, live process bubbles |
🎬 Demo video (3 minutes)
Quark drive: https://pan.quark.cn/s/4e9ebc5c07b7 ("启灵-agent")
🖼 Screenshots
| Single chat: token streaming + thinking / tool trace | Group orchestration: task-plan card |
|---|---|
 |
 |
| Live process bubble of a running task | Shared room workspace: in-browser terminal |
 |
 |
▶️ Run the demo (local)
Prerequisites: Docker / Docker Compose and mise (manages tasks plus Go / Node / pnpm / sqlc tool versions).
# 1. Install dependencies and toolchain
mise run setup
# 2. Start the dev environment (docker compose, SQLite, auto-build)
mise run dev
Then open the web console:
http://localhost:19082
The web port can be overridden via
MEMOH_SQLITE_DEV_WEB_PORT(default 19082). Stop withmise run dev:down:sqlite; list all tasks (includingdev:postgres) withmise tasks.
The demo walkthrough (single-chat token streaming, live process bubbles during group-chat orchestration, pinned long-term context, and the shared workspace file browser + in-browser terminal) is described in the Product / Technical documents.
Project overview
- Backend Go (Echo) · Frontend Vue 3 + Vite · Desktop Electron · Storage SQLite / PostgreSQL · Retrieval Qdrant
- Agent integration: memoh (built-in framework agent) / Claude Code / Codex (CLI-backed), one adapter layer + Orchestrator DAG engine, LLM + rule-based two-tier Planner
- Collaboration: AgentHub rooms, @-mentions, shared workspace, pinned long-term context, event sourcing, live in-progress process bubbles, failure degradation, and crash self-healing
Full architecture, data model, and core sequence flows are in deliverables/agenthub-technical-design.pdf.
License
See LICENSE.
Directories
¶
| Path | Synopsis |
|---|---|
|
cmd
|
|
|
agent
command
|
|
|
bridge
command
|
|
|
memoh
command
|
|
|
internal
|
|
|
agent/background
Package background implements a background task manager for long-running commands executed inside bot containers.
|
Package background implements a background task manager for long-running commands executed inside bot containers. |
|
agenthub/orchestrator
Package orchestrator implements AgentHub's task orchestration control plane.
|
Package orchestrator implements AgentHub's task orchestration control plane. |
|
agenthub/providers
Package providers implements AgentProvider adapters for external CLI-based coding agents (Claude Code, OpenAI Codex).
|
Package providers implements AgentProvider adapters for external CLI-based coding agents (Claude Code, OpenAI Codex). |
|
channel
Package channel provides a unified abstraction for multi-platform messaging channels.
|
Package channel provides a unified abstraction for multi-platform messaging channels. |
|
channel/adapters/dingtalk
Package dingtalk implements the DingTalk channel adapter.
|
Package dingtalk implements the DingTalk channel adapter. |
|
channel/adapters/feishu
Package feishu implements the Feishu (Lark) channel adapter.
|
Package feishu implements the Feishu (Lark) channel adapter. |
|
channel/adapters/feishu/wsclient
Package wsclient is a Feishu/Lark long-connection websocket client.
|
Package wsclient is a Feishu/Lark long-connection websocket client. |
|
channel/adapters/local
Package local implements the CLI and Web channel adapters for local development.
|
Package local implements the CLI and Web channel adapters for local development. |
|
channel/adapters/misskey
Package misskey implements the Misskey channel adapter.
|
Package misskey implements the Misskey channel adapter. |
|
channel/adapters/telegram
Package telegram implements the Telegram channel adapter.
|
Package telegram implements the Telegram channel adapter. |
|
channel/common
Package common provides shared utilities for channel adapters.
|
Package common provides shared utilities for channel adapters. |
|
conversation
Package conversation defines conversation domain types and rules.
|
Package conversation defines conversation domain types and rules. |
|
conversation/flow/botruntime
Package botruntime introduces a conversational runtime abstraction so a bot can be backed by different agent frameworks.
|
Package botruntime introduces a conversational runtime abstraction so a bot can be backed by different agent frameworks. |
|
memory/qdrant
Package qdrant wraps the official github.com/qdrant/go-client SDK, providing a thin facade for sparse-vector memory operations.
|
Package qdrant wraps the official github.com/qdrant/go-client SDK, providing a thin facade for sparse-vector memory operations. |
|
memory/sparse
Package sparse provides a Go client for the sparse encoding Python service.
|
Package sparse provides a Go client for the sparse encoding Python service. |
|
pipeline
Package pipeline implements the Deterministic Context Pipeline (DCP) for assembling LLM context from canonical IM events.
|
Package pipeline implements the Deterministic Context Pipeline (DCP) for assembling LLM context from canonical IM events. |
|
storage
Package storage defines the Provider interface for object storage backends.
|
Package storage defines the Provider interface for object storage backends. |
|
storage/providers/containerfs
Package containerfs implements storage.Provider for bot containers backed by gRPC calls to the in-container MCP service.
|
Package containerfs implements storage.Provider for bot containers backed by gRPC calls to the in-container MCP service. |
|
storage/providers/fallback
Package fallback implements storage.Provider that tries a primary provider first (e.g.
|
Package fallback implements storage.Provider that tries a primary provider first (e.g. |
|
storage/providers/localfs
Package localfs implements storage.Provider backed by the host filesystem.
|
Package localfs implements storage.Provider backed by the host filesystem. |
|
tui/local
Package local resolves the on-disk layout shared between the Memoh desktop Electron shell (apps/desktop) and the bundled CLI binary (cmd/memoh).
|
Package local resolves the on-disk layout shared between the Memoh desktop Electron shell (apps/desktop) and the bundled CLI binary (cmd/memoh). |
|
workspace/bridge
Package bridge provides a gRPC client for the workspace container bridge service.
|
Package bridge provides a gRPC client for the workspace container bridge service. |
|
Package spec Code generated by swaggo/swag.
|
Package spec Code generated by swaggo/swag. |
Click to show internal directories.
Click to hide internal directories.