Johnson-Durui/ApexCluster
Developer ToolsSelf-hosted multi-vendor agent cluster that turns a requirement into a PRD, prototype and Docker-verified code.
This project ships 2 agent skillsSkill list and install →Overview
ApexCluster is a locally run agent orchestration workbench for product and software engineering work. Instead of letting one model narrate its own success, it separates the job into two layers: an architecture layer that turns a rough idea into PRD.md, ARCHITECTURE.md, DEV_GUIDE.md, a CONTRACT.json acceptance contract and an interactive demo; and an implementation layer that plans, writes, cross-reviews and verifies code.
The design enforces independence between vendors. Implementation authors and reviewers must come from different model families, and stage transitions are gated by the program state machine, artifact hashes, structured command receipts and a final TRACE rather than by model text claiming "tests passed". Five flagship seats are weighted 4 → 3 → 2 and can only delegate downward; four worker pools support up to 64 concurrent tasks each. It targets product managers who want a usable PRD/prototype and engineering teams willing to configure multiple providers. Built-in presets let you start with a single provider or go full multi-vendor, and a local CLI mode can reuse already-logged-in Claude Code and Codex CLI accounts.
Getting Started
Zero-cost offline walkthrough (no API keys). Requires Node.js >= 22.12, Git and a running Docker daemon. This runs a scripted provider with no real model calls but does build and execute a real isolated Docker container:
git clone https://github.com/Johnson-Durui/ApexCluster.git
cd ApexCluster
npm ci
docker build -t dmw-sandbox:local docker/sandbox
npm run demo:offline
A successful run ends with PASS.
Real pipeline. Copy .env.example to .env and fill in provider keys, then copy config/chat.config.example.json to config/chat.config.json, set a local image's immutable sha256: ID and a budget.maxUsd cap (see the Docker execution guide). With only an OpenAI key you can exercise the architecture layer and stop at gate A:
npm run models:verify -- --provider openai
mkdir product
npm run pipeline -- --preset single-provider --requirement tests/fixtures/pipeline/todo-cli/requirement.md --workspace ./product
With all five provider keys configured, --preset flagship runs the full architecture-plus-implementation flow. The CLI emits NDJSON and exits with 0 on success, 3 when a human gate is reached in a non-interactive terminal, and 4/5 for run failures or unavailable configuration. Roster presets are documented in the configuration guide.
Web console. Run npm ci --prefix web, npm run build:web and npm start; the workbench listens only on 127.0.0.1 (default port 5170). Note that a browser session started this way cannot authorize new product directories, so use npm run pipeline when running against local directories. A no-backend preview with preset data is available as an online demo, and the delivered RoomDesk sample app can be tried directly.
Constraints
- Node.js >= 22.12, Git and Docker are required for the offline and sandboxed flows; alternative execution backends are WSL
bwrapand Windows MXC. - Real pipelines need API keys for each provider in the chosen roster; the
single-providerpreset only covers the architecture layer up to gate A. - A target workspace directory must already exist and be authorized before the pipeline can read or write it.
- Sandboxed commands run non-root with networking disabled by default; dependency installation must go through the separate controlled install channel.
- The web console binds to loopback only and deliberately does not expose key values to the frontend.
- Costs are capped per run via
budget.maxUsdandbudget.maxTokens; model IDs must match the local registry and passmodels:verify. - The project is MIT-licensed.
HotRepo's take
ApexCluster appeared on the new-project list for TypeScript on 2026-11-22, gaining 5 stars within about an hour at a total of 105 stars and 11 forks. That is a short and narrow observation window, so the numbers support only that early attention existed, not a durability judgment. What stands out technically is the emphasis on procedural gates — cross-vendor review independence, hashed artifacts, immutable container digests and a no-silent-fallback model registry — which are concrete mechanisms rather than generic "multi-agent" claims. The trade-off is setup weight: the full path expects keys for five frontier providers plus a pinned sandbox image, so the single-provider and scripted offline modes matter most for first evaluation.
Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.
You might also like
- superplanehq/superplaneOrchestrates coding agents, CI and review so routine backlog issues become verified pull requests.
- muaaz9329/cheap-agAn MCP server for Claude Code where a frontier model plans and writes tasks while a cheap model implements them under test validation.
- paperclipai/paperclipRuns a team of heterogeneous coding agents like a company: goals, org chart, budgets and audit trails in one dashboard.