The Multi-Agent Orchestration Server
ClawixeaServer is a production-grade orchestration server — not an AI wrapper. Deploy it on your own hardware, connect your LLM providers, and govern every agent run through a 22-step lifecycle, container-isolated execution, and a live multi-node operations console.
An Orchestration Server Is Not an AI Wrapper
AI API wrappers send a prompt and return text. ClawixeaServer manages the full server-side infrastructure your agents run on.
Server Infrastructure, Not an AI Plugin
ClawixeaServer provides the server-layer primitives your agents need: lifecycle management, container dispatch, job queuing, channel routing, and memory persistence.
22-Step Agent Lifecycle
Every agent run is deterministically managed server-side: load definition → build context → dispatch container → run reasoning loop → save messages → consolidate memory → record usage. Nothing happens outside the lifecycle.
- Fully managed runner service
- Cancellable at any stage
- Full run history in Postgres
Container Orchestration
Each run gets a fresh Docker container managed via CLI. A warm pool keeps containers pre-started for <50ms dispatch. Security profile: non-root 1000:1000, --network none, PID limit 256, no-new-privileges.
- Docker lifecycle via server CLI
- Warm pool — <50ms dispatch
- Hardened sandbox per run
Multi-Node High Availability
Deploy across multiple nodes. BullMQ distributes work via Redis. The stall detector re-queues failed runs within 15 seconds; a live node picks them up within 5 seconds — under 30 seconds end-to-end.
- BullMQ job queue + Redis pub/sub
- 30-second failover SLA
- Node admission via Operations Console
Governance & Audit
RBAC controls who runs which agents. Per-policy token budgets cap spend at the server level. Every tool call, token count, and status transition is logged in an immutable server-side audit trail.
- Role-based access control
- Per-policy token budgets
- Immutable audit log
- Persona-drift detection, flagged live in-chat
Channel Gateway
One server routes inbound messages from Web (WebSocket), Telegram, WhatsApp, and Slack to the correct agent session. Each channel has a typed adapter and a formatter for outbound markup.
- Web · Telegram · WhatsApp · Slack
- Channel-native formatters
- Redis fan-out for multi-node delivery
Memory Consolidation & Fine-Tuning
The memory consolidation service runs after every session, writing key facts to user-, group-, and org-scoped files. A nightly cron export closes the fine-tuning loop — interactions become training data.
- User / group / org memory scopes
- Post-run consolidation service
- Nightly JSONL fine-tune export
Industry Packs
Pre-built agents and skills for five industries. Every output is a draft — a human approves anything sent, filed, or paid.
Finance & Accounting
Bookkeeping, reconciliation, and audit — every entry and payment is human-approved.
Legal & Compliance
Contract review, case research, and drafting — always a draft, never legal advice.
NGO Operations
Donor engagement, M&E, and impact reporting — dignity-first data handling.
Home Build & Construction
Take-offs, quotes, and schedules — prices from named suppliers, never guessed.
Healthcare & Clinical Ops
Documentation, coding, and claims — every note awaits physician sign-off.
How the Orchestration Server Handles a Run
Every request follows the same deterministic server-side lifecycle — observable, cancellable, and safe to interrupt at any stage.
Channel gateway receives message
Web WebSocket, Telegram, WhatsApp, or Slack adapter normalises the inbound message to the internal format and routes it to the session resolver.
Agent-runner builds context
Server loads agent definition, assembles system prompt from memory files, wiki context, and recent session history.
Container pool dispatches
A pre-warmed Docker container is assigned from the pool in under 50ms. Non-root, network-isolated, PID-limited — a clean sandbox every time.
Reasoning loop executes
LLM ↔ tool calls iterate inside the container up to 40 rounds. BullMQ writes a heartbeat every tick — live status visible in the Operations Console.
Memory consolidation service runs
Key facts from the session are written back to user-, group-, and org-scoped memory files so the next run starts with full context.
Audit record committed
Run status, tool calls, token usage, worker node, and duration are persisted to Postgres. No run leaves the server without a complete record.
Server Code You Can Read, Audit, and Extend
ClawixeaServer is MIT-licensed. The container hardening, RBAC rules, channel adapters, job queue wiring, and 22-step agent lifecycle are all in plain TypeScript — no proprietary runtime, no black-box orchestration layer.
Frequently Asked Questions
Deploy the Orchestration Server in 5 Minutes
Clone the repo, run pnpm run install:clawixserver, and your orchestration server is live — Postgres, Redis, Docker agent sandbox, and the full channel gateway configured automatically.
Or start with 2 × DigitalOcean droplets at ~$284/mo for full HA.