🌐 Languages: English · 中文 · 日本語 · 한국어 · Español · Français · Deutsch
🔥 Hot (2026-07): DeepSeek-V4-Flash-0731 official release — agentic capability sharply upgraded, beating the pricier GLM-5.2 on agent benchmarks (Terminal Bench 82.7 vs 81.0, DeepSWE 54.4 vs 46.2, Toolathlon 70.3 vs 59.9) at far lower cost. MoA's Flash tier (tool + opinion layers) got much stronger at the same low price.
🔄 LongLoop (24h unattended): Keep your project iterating for days — no forgetting, no stopping, no repeating. The concierge wakes every round, routes through the full MoA pipeline, and writes progress to disk. One command to run it on your project: ▶ Get started →
One conversation entry point, 22 specialized models collaborating automatically. Simple tasks use Flash (cheap), complex tasks call the flagship (expensive). Cost down up to ~90% (vs all-flagship) when simple tasks dominate the workload and flagship calls are minimized — actual savings depend on task mix; code quality significantly up.
structured output (---section--- markers), 界线 acceptance, anti-cheat, auto-routing. See CHANGELOG.
OpenCode MoA is a Mixture of Agents configuration package for OpenCode. It lets multiple models think about the same problem simultaneously, then fuse into an output quality a single model can't reach. You don't need to switch tools, write code, or have an API quota — just drop the files into your project and restart OpenCode.
22 agents · 5 commands · 3 skills · 30-second deploy
By default OpenCode uses a single model from start to finish. Changing one character and designing a system architecture use the same prompt, same temperature, same context. No division of labor.
Three problems: ① cost out of control — simple tasks also use the expensive model; ② quality bottleneck — a single model has only one way of thinking; ③ no fault tolerance — if the model dies it freezes, no fallback.
MoA's solution:
You: help me design a message queue solution
┌─ flag-arch (Qwen3.7 Max) ─── plan from the architect's view
├─ flag-plan (DeepSeek V4 Flash) ─── plan from the planning view
├─ flag-eng (DeepSeek V4 Flash) ─── plan from the implementer's view
└─ flag-fuse (Kimi K3 ) ─── take the best of each, one optimal solution
Three independent plans from three different models naturally form a "consensus + divergence" structure. The fusion model keeps the consensus and takes the best where they diverge — something a single model cannot do.
| Requirement | Notes |
|---|---|
| OpenCode >= 1.3.4 | agent-level reasoningEffort/hidden/task support, install |
| OpenCode Go plan | Subscribe, first month $5, then $10/month |
| Git | used to clone the repo |
PowerShell 7+ (pwsh) |
drives LongLoop (long-loop.ps1) & install.ps1 — Windows ships 5.1, install PS7 (winget install Microsoft.PowerShell, or auto-MSI: powershell -File install.ps1 -InstallPwsh), macOS brew install powershell, Linux see LongLoop docs |
| Node.js >= 14 | runs the moa-loop MCP server (longloop/server.js) — nodejs.org |
install.sh additionally needs jq (Linux/macOS) — without them use Method 1 or Method 3 below.
⚠️ Key path pitfall — put the provider + key in either the project-levelopencode.jsonor the system-level shared path, pick one. System-level correct path: Linux/macOS~/.config/opencode/opencode.json; Windows%USERPROFILE%\.config\opencode\opencode.json(not%APPDATA%\opencode). Wrong path → "deployment succeeds but all agents can't connect".
- Clone the repo:
git clone https://github.com/ZenHG/opencode-moa.git - In OpenCode (opened in your project), send:
Deploy all 22 agents, 5 commands, and 3 skills from the opencode-moa repo into the current project
- The AI reads the real source files (
.opencode/agents/,opencode.json) and creates all files automatically. Restart OpenCode when done.
The repo itself is the installer — the AI deploys from the actual source files, so there is no manual to download or keep in sync.
# clone the repo, then copy the config into your project
git clone https://github.com/ZenHG/opencode-moa.git
cd your-project
cp -r ../opencode-moa/.opencode/ .
cp -r ../opencode-moa/.moa/ .
cp -r ../opencode-moa/longloop/ .
# run the install script (auto-merge config, keeps your API key)
# Windows: pwsh ../opencode-moa/install.ps1 (auto-installs missing deps: -InstallPwsh / -InstallNode / -InstallDeps for all)
# Linux/macOS: bash ../opencode-moa/install.sh --install-jq (auto-installs jq if missing)The install script auto-backs up your original
opencode.json, only merging MoA config while keeping your provider and API key.
# 1. clone the repo
# 2. copy the .opencode directory and .moa config into your project
# 3. manually merge opencode.json (do NOT replace directly!)
# merge MoA's permission.task and agent sections in, keep your existing provider and model config
⚠️ Do not usecat >>to append (corrupts JSON) and do not replace directly (loses your API key).
MoA is a generic template — every agent's model is just an ID you can change. Each agent file starts with model: opencode-go/<model-id>. Swap a model by editing that one line in .opencode/agents/<agent>.md (e.g. opencode-go/kimi-k2.7-code, opencode-go/deepseek-v4-flash). No reinstall needed.
- Restart OpenCode, press
Tabto cycle agents (Windows desktop:Ctrl+.also works) and see 门童 - Type
@工具人and it responds - Run
pwsh .opencode/tests/T0-static-verify.ps1— expected all PASS
rm -rf your-project/.opencode/
rm -rf your-project/.moa/
# manually restore your opencode.json (the install script auto-backs up a .bak file)Learn nothing — just talk. 门童 (concierge-router) automatically judges task complexity and dispatches the corresponding agent chain.
| What you say | What 门童 does | Agents used |
|---|---|---|
| "rename this variable" | judged as a simple task | swift (Flash) |
| "write a user auth module" | tool layer gathers → 3 mid-tier parallel → fuse | tool-handler + mid-tier trio + fuse |
| "design a microservice architecture" | tool layer gathers → 3 flagship parallel → fuse → implement → QA | full-chain 6 agents |
| "restore this screenshot's UI" | 3 frontend experts parallel → lead picks best | frontend quartet |
| message with screenshot | vision-translator converts to text → normal routing | vision-translator |
| message with error log / diagram / complex content | vision-translator decomposes content → normal routing | vision-translator (fallback role) |
Direct @ calls (visible agents only): @闪电侠 help me write a hello world · @工具人 search all TODOs in the project · @视觉翻译 analyze this screenshot — the other 18 agents are hidden from the @ menu; 门童 calls them via the Task tool automatically.
One-click commands:
| Command | Scenario |
|---|---|
/moa-quick |
simple task, translation, config change |
/moa-medium |
function module, bug fix, single-file refactor |
/moa-flagship |
system architecture, large refactor |
/moa-frontend |
UI restore, CSS, screenshot fix |
/moa-describe |
screenshot/image to text |
门童 auto-detects task type via keyword analysis: exploration tasks ("analyze", "compare", "understand", "investigate") → exploration prompt + exploration acceptance specs; execution tasks ("fix", "add", "implement", "deploy") → execution prompt + stop-loss rules. The task type (taskType=explore|execute) is inlined as metadata to the fusion layer, which generates the matching acceptance criteria.
concierge-router (Flash)
│
┌────────────────┼─────────────────┐
▼ ▼ ▼
Tool layer Opinion layer Fusion layer
Flash + MiMo 3 parallel opinions take the best
(~80% calls) (~18% calls) (~2% calls)
- Tool layer (Flash + MiMo + Qwen3.7 Plus) — read code, search files, screenshot to text. Cheap and fast, call freely.
- Opinion layer (Qwen / Kimi / Flash) — plans from different perspectives; three opinions naturally form "consensus + divergence".
- Fusion layer (Kimi K3 / Kimi K2.7 / Flash lead / DeepSeek V4 Pro fallback) — keep consensus, take the best on divergence. The flagship fuse runs on Kimi K3 (2.8T params, 1M context) — MoA's quality ceiling is at the front of the pack.
⚠️ The call-volume ratios (~80% / ~18% / ~2%) are design targets, not measured statistics. Actual ratios vary by task complexity.
Opinion and fusion agents use ---section-name--- markers for structured output (opinion: ---记忆层--- + ---方案--- + ---红线---; fusion: ---融合方案--- + ---分歧裁决--- + ---白名单--- + ---红线--- + ---验收标准---), enabling downstream parsing and acceptance verification. Anti-cheat prevents implementation agents from cutting corners: baseline non-regression, forbidden actions (skip/mock/delete tests), hidden spot-checks (暗线), stop-loss. Acceptance criteria are frozen in .moa/界线.json. Deep-dive: fault tolerance, cost, security, FAQ.
The English name is the logical role; the Chinese in parentheses is the exact filename under
.opencode/agents/. Call visible agents directly with@(@门童,@工具人,@闪电侠,@视觉翻译); the 18[hidden]agents are orchestrated by 门童 via the Task tool and are not directly @-callable.
concierge-router (门童, Flash)
│
├── Tool layer ─────────────────────────────────────────────
│ tool-handler (工具人, Flash ) read code, search files
│ tool-handler-mimo (工具人-mimo, MiMo) [hidden] reliable file read (fallback + parallel)
│ swift (闪电侠, Flash ) simple tasks in one shot
│ vision-translator (视觉翻译, Qwen3.7 Plus ) screenshot/UI→text; logs/diagrams/docs→decomposition
│
├── residual-extractor (残差提取, Flash ) analyze divergence between plans
├── confidence-assessor (置信度评估, DeepSeek V4 Flash ) assess fusion result confidence
│
├── Mid-tier opinion layer ─────────────────────────────────────────────
│ mid-eng (中级·工程, Kimi K2.6 ) engineering view
│ mid-creative (中级·创意, Qwen3.7 Plus) creative view
│ mid-coder (中级·码农, Flash ) pragmatic view
│ mid-fuse (中级·融合, Kimi K2.7 Code) fuse three plans [max_tokens: 16384]
│
├── Flagship opinion layer ─────────────────────────────────────────────
│ flag-arch (旗舰·架构, Qwen3.7 Max ) top-level architecture
│ flag-plan (旗舰·规划, DeepSeek V4 Flash) structured planning
│ flag-eng (旗舰·工程, DeepSeek V4 Flash ) large-scale implementation
│ flag-fuse (旗舰·融合, Kimi K3 ) fuse three architecture plans [max_tokens: 16384]
│ flag-impl (旗舰·执行, Flash) [hidden] implement per fused plan
│ flag-qa (旗舰·质检, DeepSeek V4 Pro) plan review + code acceptance [max_tokens: 16384]
│
└── Frontend opinion layer ─────────────────────────────────────────────
fe-restore (前端·还原, Qwen3.7 Plus ) pixel-perfect UI restore
fe-logic (前端·逻辑, Qwen3.7 Plus) component architecture & state mgmt
fe-motion (前端·动效, MiMo-Pro ) interaction & motion
fe-lead (前端·总工, DeepSeek V4 Flash) pick best of three frontend plans [max_tokens: 16384]
Fallback agent (not in the router chain above, called only when fusion fails):
fallback (融合·保底, DeepSeek V4 Pro) — same residual-enhanced fusion, used when flag-fuse / mid-fuse / fe-lead fail
| Doc | Contents |
|---|---|
| docs/README-details.md | Fault tolerance design · cost model · security · local models · verification · FAQ · maintainer tooling |
# Layer 0 — static check (automatic, 0 token)
pwsh .opencode/tests/T0-static-verify.ps1
# run all three layers at once
pwsh .opencode/tests/run-all.ps1Check scripts under .opencode/tests/: Layer 0 automatic (T0 static / T1 README consistency / T3 permission security); Layers 1–2 guided checklists. Details: Verification.
PRs and Issues welcome. See CONTRIBUTING.md.