Key takeaways
- 16,000+ GitHub stars as of June 2026 (created February 23, 2026) — one of the fastest-growing Chinese-origin agent orchestration projects
- The 门下省 Gate Review layer is an institutional veto: a dedicated agent audits every plan and can hard-reject it before execution, a review primitive most agent frameworks lack
- Free and MIT licensed, self-hosted on OpenClaw, with a real-time "War Cabinet" dashboard for kanban, agent health, token spend, and per-agent model hot-switching
FAQ
What is Edict?
Edict is an open-source, self-hosted multi-agent orchestration system built on OpenClaw that organizes AI agents as officials in the Tang Dynasty's Three Chancelleries and Six Ministries bureaucracy, with a mandatory review agent that can reject plans before execution. MIT licensed.
How much does Edict cost?
The software is free (MIT, self-hosted). You pay for the LLM API tokens your agents consume — and the mandatory review layer adds extra model calls per task.
How does the Gate Review (门下省) work?
The planning agent (中书省) decomposes a directive into an execution plan; the review agent (门下省) audits that plan for quality and completeness and can approve it or hard-reject it, forcing re-planning. The checkpoint cannot be bypassed.
How is Edict different from Paperclip?
Paperclip models agents as employees in a Western-style org chart with budgets and heartbeats; Edict models them as imperial officials with separation of powers, and its differentiator is the institutional veto layer that reviews plans before anything executes.
Executive Summary
Edict is an open-source multi-agent orchestration system that organizes AI agents using the governance model that ran imperial China for over a millennium: the Tang Dynasty's 三省六部制 (Three Chancelleries and Six Ministries). Directives flow from the human "Emperor" through a triage agent, a planning chancellery, a review chancellery, and a dispatch chancellery to six specialized ministry agents — with full audit trails and a real-time dashboard.[1] The repository describes the system as nine specialized agents; the README documents an extended roster including a 太子 (Crown Prince) triage agent and a 早朝官 daily-briefing agent.[2]
The killer feature is the 门下省 (Gate Review) layer: a dedicated agent whose only job is to audit plans from the planning chancellery and veto them — a hard reject that forces re-planning — before anything reaches execution. The checkpoint cannot be bypassed, which the project positions as its core difference from frameworks like CrewAI and AutoGen that lack a systematic review layer.[1] The project is Chinese-language-first, built on OpenClaw, and has grown to 16,000+ stars on GitHub as of June 2026.[3]
| Attribute | Detail |
|---|---|
| Creator | cft0808 (GitHub handle; no disclosed company) |
| Created | February 23, 2026 [3] |
| Funding | None disclosed |
| GitHub Stars | 16,037 (June 2026) [3] |
| Forks | 1,687 [3] |
| License | MIT [3] |
| Primary Language | Python [3] |
| Docs | Chinese primary; English and Japanese READMEs [1] |
Product Overview
You install Edict on top of OpenClaw, define your agents' workspaces and permission matrices via an install script, and issue directives ("edicts") as the Emperor. The 太子 agent triages whether a message is casual chat or an official directive; official directives get planned by 中书省, reviewed (and possibly rejected) by 门下省, dispatched by 尚书省 to the appropriate ministry agents, and reported back — every stage logged.[1]
The six ministries map classical portfolios to modern work: 吏部 (personnel/agent management), 户部 (revenue/data), 礼部 (rites/documentation), 兵部 (war/engineering), 刑部 (justice/compliance), and 工部 (works/infrastructure).[1]
Key Capabilities
| Capability | Description |
|---|---|
| Gate Review (门下省) | Mandatory plan audit with veto power; rejection forces re-planning [1] |
| Message triage (太子) | Separates casual chat from official directives before the pipeline engages [1] |
| War Cabinet dashboard (军机处) | 10 panels: kanban with task intervention, agent health heartbeats, memorial archive with five-stage timelines, template library, performance rankings, token metrics, news aggregation with Feishu integration, skills management, session monitoring, multi-agent court debate [1] |
| Model hot-switching | Per-agent model configuration changes take effect in ~5 seconds without restart [1] |
| Audit trails | Full state-machine validation; illegal status transitions are blocked [1] |
Product Surfaces
| Surface | Description | Availability |
|---|---|---|
| Self-hosted dashboard | React web UI on port 7891 | Open source [1] |
| Demo container | Preloaded demo data via a single container run | Open source [1] |
| systemd service | Daemon deployment with auto-restart | Open source [1] |
Technical Architecture
Edict deploys self-hosted only. The quickest path is the Docker demo image:
docker run -p 7891:7891 cft0808/sansheng-demo
Full installation clones the repo and runs install.sh, which configures agent workspaces, SOUL.md persona files, permission matrices, and symlinked shared data; a single command then starts the dashboard API and data refresh loops (15-second sync intervals).[1]
Key Technical Details
| Detail | Value |
|---|---|
| Deployment | Self-hosted (Docker demo, install script, systemd) [1] |
| Model(s) | Multi-provider, hot-swappable per agent [1] |
| Agent runtime | Built on OpenClaw [2] |
| Backend | Python stdlib http.server, Redis EventBus with Outbox Relay pattern, SQLAlchemy [1] |
| Frontend | React 18 + TypeScript + Zustand [1] |
| Open Source | Yes — MIT [3] |
Strengths
- Institutional veto as a primitive — the 门下省 Gate Review is a structural answer to unvetted agent output: plans get audited and can be hard-rejected before execution, by design rather than by prompt.[1]
- Rapid adoption — 16,037 stars in under four months (created February 23, 2026) with active pushes as recently as June 4, 2026.[3]
- Deep observability — a 10-panel dashboard covering kanban, agent heartbeats, token consumption rankings, and per-task five-stage timelines, plus state-machine enforcement of status transitions.[1]
- Lightweight stack — no external backend dependencies beyond Redis; backend runs on the Python standard library.[1]
- Strict permission matrix — Chinese community reviewers single out the permission matrix layered on OpenClaw's messaging as the project's strongest engineering, calling the institutional design "creative with real depth."[4]
- Free and auditable — MIT licensed with no paid tier, account, or telemetry requirement.[3]
Cautions
- Review latency is the price of the veto — the README itself frames the tradeoff as auditability and control over autonomous speed: every task pays for a mandatory review pass (extra model calls, extra wall-clock time) whether it needs one or not.[1]
- No stable release yet — open issues include users asking when an official version will ship, alongside recurring setup and configuration problems (console not starting, model config not applying), many sitting in a "needs-triage" backlog.[5]
- Inherits OpenClaw's security posture — V2EX users testing OpenClaw report low default security and prompt-injection risk, commonly isolating it on sandboxed machines; Edict runs on top of that substrate.[6]
- Chinese-first project — documentation, dashboard concepts, and the majority of issue traffic are in Chinese; English and Japanese READMEs exist but the community center of gravity is Chinese dev forums.[5]
- Clone confusion — the project has spawned copycat repositories, to the point that a popular Linux.do thread is titled "oppose piracy, support the original."[7]
- Agent count ambiguity — the repo description says nine specialized agents while the README documents additional support roles; the roster is still settling.[2]
- No disclosed company or funding — sustainability rests on an individual maintainer's continued interest.
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open source | Free | Full system: agents, dashboard, audit trails, templates [3] |
Licensing model: MIT — permissive, self-hosted, no account required.[3]
Hidden costs: LLM API tokens for every agent in the pipeline — the mandatory Gate Review pass adds model calls to every task — plus Redis hosting and the operational overhead of a self-hosted deployment.[1]
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| Paperclip | Western org-chart metaphor (CEO/CTO, budgets, heartbeats) and agent-agnostic across all business functions; governance is human-approval-based. Edict's governance is agent-on-agent — an institutional review layer vetoes plans before a human ever needs to. |
| CrewAI / AutoGen | General multi-agent frameworks; Edict positions its mandatory, non-bypassable review checkpoint as the difference from their freer agent topologies [1] |
| OpenClaw (bare) | Edict's own substrate — OpenClaw alone routes by @mention and subagent spawning; Edict adds the bureaucratic pipeline, permission matrix, and dashboard on top [4] |
When to Choose Edict Over Alternatives
- You already run OpenClaw and want structured governance instead of free-form agent chatter
- Unvetted agent output is your top failure mode — you want plans rejected before execution, not flagged after
- You read Chinese, or can tolerate Chinese-first docs and community
- You want deep per-task audit timelines and token accounting out of the box
Ideal Customer Profile
Best fit:
- OpenClaw power users (especially in Chinese-speaking dev communities) who want institutional structure over their agents
- Teams whose tasks are high-stakes enough that mandatory plan review is worth the latency
- Tinkerers drawn to the historical-systems framing who want a self-hosted, auditable control plane
Poor fit:
- Teams that need low-latency, high-throughput agent execution — every task pays the review tax
- English-only enterprises wanting vendor support, SLAs, or a stable release
- Anyone unwilling to sandbox an OpenClaw deployment given its known prompt-injection exposure [6]
Viability Assessment
| Dimension | Assessment |
|---|---|
| Financial Health | No company or funding; free OSS sustained by an individual maintainer |
| Market Position | Standout Chinese-origin entry in agent-company platforms; 16k+ stars in under four months [3] |
| Innovation Pace | Active — pushed June 4, 2026; 33+ open issues with steady feature requests [5] |
| Community/Ecosystem | Strong in Chinese dev forums (Zhihu, Linux.do); essentially absent from Hacker News and English-language Reddit [7] |
| Long-term Outlook | Depends on a stable release and maintainer bandwidth; the Gate Review idea is likely to outlive the implementation |
Edict's trajectory looks like a classic single-maintainer breakout: explosive star growth, a genuinely novel governance primitive, and a triage backlog that hasn't caught up. The idea is stronger than the operational maturity — which is normal at three and a half months old.
Bottom Line
Edict is the most interesting governance experiment in the agent-company category: instead of asking a human to approve everything (or trusting agents to self-correct), it institutionalizes dissent — a dedicated agent whose job is to say no before execution. The Tang Dynasty framing is more than aesthetic; separation of powers maps surprisingly well onto multi-agent failure modes. But it is Chinese-first, pre-1.0, built on a substrate with known security caveats, and maintained by one pseudonymous developer.
Recommended for: OpenClaw users who want structured, auditable agent governance and accept review latency as the cost of vetoed mistakes.
Not recommended for: English-only teams needing a supported, stable platform, or workloads where execution speed beats plan quality.
Outlook: Watch for an official release and English-community traction; the Gate Review pattern is the export — expect other platforms to copy it.
Research by Ry Walker Research • methodology