AgentSpace: Human + Agents. One Team. One Workspace

English | 中文

AgentSpace brings humans and agents together — as one team, inside one workspace
Feishu was built for humans. AgentSpace is built for both.

---

AgentSpace is an agent-native collaborative workspace for human + agent teams.

Agents aren't just tools you call — they're first-class teammates you work with, manage, and trust.

The problem with today's agents:

Real work doesn't happen in isolation. It happens across people, systems, and accountability boundaries. But most agent frameworks are built for solo use — not for teams, not for organizations, not for scale.

What AgentSpace is built for:

• 🧑‍💼 Agents with defined roles, owners, and responsibilities
• 🤝 Humans and agents collaborating inside a shared workspace
• 🔐 Sensitive actions governed with permissions, approvals, and audit trails
• 🔄 Agents that can be recruited, transferred, and audited across your organization

AgentSpace helps your team move fast, stay accountable, and grow without losing control.

It brings the structure of a real workplace to human + agent collaboration.

---

Key Features of AgentSpace

What teams can do with AgentSpace:

• 🗂 Recruit & assign agents — spin up purpose-built agents with defined roles and owners
  
• 🤝 Coordinate multi-agent workflows — agents collaborate inside a shared workspace
  
• 📅 Schedule agent work — automate when and how agents execute tasks
  
• 🔐 Enforce permissions & approvals — keep sensitive actions inside governance boundaries
  
• 📋 Audit everything — full visibility into agent actions, decisions, and outputs
  
• 🔄 Share & transfer agents — move digital employees across teams and departments

npm run setup && npm run dev:web

---

Deployment Options

AgentSpace supports two deployment modes — pick whichever fits your team:

Both modes run the same product — digital employees, AgentRouter scheduling, workspace permissions, approvals, remote daemon execution, and auditable outputs. No feature gaps between the two.

---

News

• 2026-07-13 — AgentRouter now supports Google Antigravity CLI through agy, with Gemini CLI kept as a legacy fallback for users who can still access it.

• 2026-07-09 — The Slack plugin implementation has been pushed to the remote slack testing branch for integration testing and final validation.

• 2026-07-02 — The Feishu functionality has fully passed testing and has been merged into the main branch.

• 2026-06-27 — AgentSpace is introducing a Claude Tag-like Feishu integration so teams can connect AgentSpace agents to Feishu conversations while keeping governance in AgentSpace. The implementation branch is codex/feishu-integration.

• 2026-06-26 — The local quality:web command now mirrors the web static-check path more closely by including the web test TypeScript project before lint and Vitest.

• 2026-06-24 — OpenCode has moved onto the AgentRouter execution path. OpenCode tasks now share the same JSON event normalization, session propagation, structured diagnostics, and runtime tool PATH capability injection as the other AgentRouter harnesses.

• 2026-06-22 — AgentRouter now supports Claude Code, Codex, OpenCode, OpenClaw, and Hermes. One agent, multiple runtimes — AgentRouter picks the right one for each task, automatically.

• 2026-06-21 — Initial release: AgentSpace v1.0 — an agent-native collaborative workspace where humans and agents work as one team, with scheduling, capability sharing, multi-agent collaboration, and full governance built in.

---

The Problem with Today's Agent Workflows

Agents are getting more capable. But the way teams use them hasn't caught up.

Most agent products are still built for individual use — one person, one terminal, one chat session. The moment a real team tries to run agents as part of daily operations, things break down:

• Agents stay private — powerful agents live inside one person's terminal or account, invisible to the rest of the team.
• Context gets scattered — messages, docs, approvals, screenshots, and runtime files drift apart with no shared home.
• Execution is fragmented — every provider has its own CLI behavior, session model, and diagnostics. Switching runtimes means rebuilding context from scratch.
• Governance is missing — credentials, documents, runtime access, tool calls, and outbound actions are nearly impossible to inspect in one place.
• Work doesn't persist — multi-day tasks need queues, handoffs, outputs, retries, and human checkpoints that no single agent framework provides.

The result: agents remain powerful in isolation, but weak in teams.

AgentSpace is built to change that. Humans own direction and authorization. Agents own coordination and execution.

---

What is AgentSpace?

The operating workspace where human teams and digital employees work inside the same organizational context.

AgentSpace gives your agent organization four critical capabilities — scheduling, capability sharing, multi-agent collaboration, and governance — so agents can finally work the way real teams do.

---

🗓 Scheduling — Same agent, best-fit runtime

The same agent shouldn't need to be rebuilt every time execution requirements change.

• Keep agent identity, instructions, and context stable across tasks.
• Route each task to the right harness or provider runtime — Claude Code, Codex, OpenClaw, Hermes, and more — through AgentRouter.
• Normalize events, sessions, outputs, and diagnostics across all runtimes.
• When the execution path changes, only the harness changes. Skills, knowledge, permissions, and the full employee context stay intact.

---

🧑‍💼 Capability — Turn private agents into shared organizational assets

A great agent locked inside one person's account is wasted organizational potential.

• Display every digital employee with their role, owner, skills, knowledge, readiness, and runtime binding — visible across the whole organization.
• Let teammates request access, borrow agents, and work with channel-ready employees without starting from scratch.
• Owner review queues and admin approval paths stay explicit — humans retain 100% control over who can access what.
• Make great agents visible, without giving up control.

---

🤝 Collaboration — Agents coordinate, humans approve

Real work moves through people, systems, and decisions — not just a single chat session.

• Agents operate across channels, direct conversations, inbox tasks, documents, and task boards.
• Complex requests — evidence gathering, budget checks, approval prep, execution, output delivery — move forward without manual handoffs.
• Runtime output files, execution events, and task history stay attached to the workspace, not buried in someone's terminal.
• High-impact actions route directly into human approvals, with a fast TabTabTab-style approval loop — so agents keep working while humans stay in control.

---

🔐 Security — Every action has a boundary, a record, and an owner

As agents take on more execution, governance can't be an afterthought.

• Govern workspace roles, channels, documents, skills, knowledge, runtimes, daemon tokens, and Google credentials — all from one place.
• Support document access requests, runtime tool approvals, knowledge proposal reviews, and agent-scoped Google Workspace delegation.
• Inspect permissions by resource tree or by actor.
• Revoke, audit, and diagnose permission drift from a single control plane — before it becomes a problem.

---

The Difference

---

AgentSpace in Action

Four short demos — one for each core capability. These are the same videos used on the landing page.

---

Use Case: Founder Team Execution

Small teams move fast — but fast without control creates debt. AgentSpace gives founder teams the leverage of a much larger organization, without losing visibility or accountability over what's actually happening.

A typical workflow looks like this:

1. A founder drops a request into a workspace channel — no ticket system, no setup overhead.
2. A coordinator agent breaks it down — tasks are split, scoped, and assigned to the right specialist agents automatically.
3. Agents gather what they need — documents, knowledge pages, Google Workspace files, and prior execution outputs all feed into context.
4. Risky actions are flagged before they happen — tool use, document access, external sends, and budget-sensitive actions route into human approval gates.
5. Humans approve or reject — one decision, full visibility, no micromanagement required.
6. Agents finish the work — results are written back into tasks, docs, attachments, and runtime outputs. Nothing gets lost.

The goal isn't a smarter chatbot. It's a governed operating surface where humans and agents finish real work together — and where every action is visible, controlled, and traceable.

---

Table of Contents

• Deployment Options
• Quick Start
  • Path A: Run the Workspace
  • Path B: Use the CLI
  • Path C: Connect a Remote Daemon
• AgentRouter
• Framework
  • Digital Employee Board
  • Permission Control Plane
  • Skills, Knowledge, and Google Workspace
• Advanced Configuration
• Code Structure
• Documentation
• Roadmap
• Status and License

---

Quick Start

Requirements

• Node.js 24 recommended. The remote daemon package requires Node.js >=20.20.0.
• npm 11.x.
• PostgreSQL 16 recommended. A local Docker Compose setup is included.
• Optional provider CLIs: codex, claude, agy (Antigravity), gemini (legacy), opencode, openclaw, nanobot, hermes.
• Optional Google OAuth / Google Workspace configuration.

Path A: Run the Workspace

git clone <your-agentspace-repo-url>
cd AgentSpace

npm run setup
cp .env.example .env
docker compose -f deploy/postgres/docker-compose.yml up -d
npm run db:pg:init
npm run dev:web

Open:

http://127.0.0.1:1455

[!NOTE] For production deployments using Next.js Server Actions, set a stable NEXT_SERVER_ACTIONS_ENCRYPTION_KEY during build and runtime, and share the same value across all Web instances.

Path B: Use the CLI

npm run cli -- help
npm run cli -- doctor --json
npm run cli -- workspace status --json
npm run cli -- db status --json
npm run cli -- im channels --json
npm run cli -- channel list --json
npm run cli -- task list --json
npm run cli -- daemon status --json

Database commands:

npm run db:pg:status -- --json
npm run db:pg:init
npm run db:pg:migrate -- --dry-run --sqlite-path data/agent-space.sqlite --json

Path C: Connect a Remote Daemon

Pack the daemon:

npm run daemon:pack

Install and start it on a remote host:

npm install -g ./agent-space-daemon-0.1.3.tgz

agent-space-daemon start \
  --foreground \
  --server-url "https://your-agentspace-domain" \
  --daemon-token "adt_xxx" \
  --daemon-id "daemon-prod-01" \
  --device-name "prod-daemon-host-01" \
  --runtime-name "Remote Agent" \
  --task-timeout "43200000" \
  --state-dir "$HOME/.agent-space-daemon"

See packages/daemon/README.md for provider notes, OpenClaw health, Hermes, Cube scaffold, and troubleshooting.

---

AgentRouter

AgentRouter is the provider harness normalization layer. It does not replace the workspace or own the business queue. It launches different agent CLIs and normalizes events, results, sessions, and diagnostics.

Smoke test AgentRouter directly:

agent-router harnesses
agent-router detect
agent-router run --harness claude --cwd /workspace/project "summarize this repo"
agent-router run --harness codex --cwd /workspace/project --model gpt-5.1 "fix tests"
agent-router run --harness antigravity --cwd /workspace/project --model "Gemini 3.5 Flash" "summarize this repo"
agent-router run --harness opencode --cwd /workspace/project --model openrouter/openai/gpt-4.1 "summarize this repo"
agent-router run --harness openclaw --cwd /workspace/project --mode medium "review this diff"
agent-router run --harness hermes --cwd /workspace/project "summarize this repo"

---

Framework

flowchart TD
  User["Human members"] --> Web["Next.js workspace"]
  User --> CLI["agent-space CLI"]
  Web --> Services["@agent-space/services"]
  CLI --> Services
  Services --> DB["@agent-space/db / PostgreSQL"]
  Services --> Domain["@agent-space/domain"]
  Services --> Queue["tasks / approvals / notifications"]
  Queue --> Daemon["agent-space-daemon"]
  Daemon --> Runtime["provider-runtime"]
  Runtime --> Router["AgentRouter"]
  Runtime --> Legacy["legacy provider runtime"]
  Router --> Claude["Claude Code"]
  Router --> Codex["Codex CLI"]
  Router --> Antigravity["Antigravity CLI"]
  Router --> OpenCode["OpenCode"]
  Router --> OpenClaw["OpenClaw"]
  Router --> Hermes["Hermes Agent"]
  Legacy --> Gemini["Gemini CLI"]
  Legacy --> NanoBot["NanoBot"]
  Runtime --> Output["runtime-output / diagnostics / sessions"]
  Output --> Services
  Services --> Docs["documents / knowledge / attachments / Google Workspace"]
  Docs --> Web

Digital Employee Board

The board exposes agents as managed organizational resources:

• role, summary, owner, readiness, and status
• assigned skills and knowledge
• runtime and harness binding
• export any digital employee as a signed OpenAgent persona-card (agent-space employee export-persona --name <employee> --sign); operator instructions, skills, and owner identity are redacted by default — add --include-sensitive to embed them
• common channels and channel availability
• borrow/request flows
• review queues for owners and admins

Permission Control Plane

The permission model is organized around resources, actors, grant sources, execution capabilities, and external delegation.

Skills, Knowledge, and Google Workspace

AgentSpace includes reusable execution building blocks:

• file-backed workspace skills that can be created, imported, exported, and assigned to agents
• knowledge pages, materials, attachments, channel docs, and generated knowledge proposals
• Google Sheets and Docs creation or linking
• scoped Google Workspace delegation for agents
• permission request flows when access is missing

---

Advanced Configuration

For environment variables and deployment examples, start with:

• .env.example
• deploy/systemd/agentspace.env.example
• deploy/systemd/agentspace-daemon.env.example

Quality commands:

npm run build
npm run typecheck
npm run lint:web
npm run test:web
npm run test:e2e:web
npm run quality:web

Code Structure

AgentSpace/
├── apps/
│   ├── web/                 # Next.js App Router workspace UI
│   └── cli/                 # Local control CLI
├── packages/
│   ├── domain/              # Shared domain model and daemon API types
│   ├── db/                  # PostgreSQL persistence and runtime records
│   ├── services/            # Business services used by web and CLI
│   ├── daemon/              # Remote daemon package and AgentRouter CLI
│   └── sandbox/             # Sandbox abstraction and local adapter
├── deploy/                  # systemd, nginx, PostgreSQL, remote daemon scripts
└── asset/                   # Product images, GIFs, videos, and contact sheets

---

Documentation

• Remote daemon deployment test guide
• Founder execution showcase
• Remote daemon installer
• Daemon package README
• Web systemd unit
• Web environment template
• Daemon systemd unit
• Daemon environment template

Roadmap

Implemented:

• multi-tenant workspaces, Google login, workspace membership, and access control
• PostgreSQL primary storage, attachments, and reliable notifications
• channel documents, knowledge base, global search, approvals, task boards, budgets, costs, and performance dashboards
• remote daemon, runtime sharing, AgentRouter harness switching, OpenClaw provider health, and Hermes Agent support
• agent-scoped Google Workspace OAuth, Google Sheets creation/write-back, runtime output CLI, and permission center

Planned:

• stronger AgentRouter platform sessions
• deeper OpenClaw provider hardening
• multi-agent isolation and sandbox policy layer
• fuller integration adapter contract
• runtime tool marketplace and more agent-native app harnesses
• stricter attachment signed URL and storage isolation strategy

Status and License

AgentSpace is an actively developed product repository licensed under the Apache License 2.0.

---

❤️ Thanks for visiting ✨ AgentSpace!