One gateway. Many agents. Distributed under a single token surface.
RedClaw is a self-hosted AI agent platform: an OpenAI-compatible gateway in front of 20+ LLM providers, a Postgres-backed agent registry, 3-tier memory with RAG, tools, teams, and multi-channel delivery. The same engine powers redxtrm, jobXlaw and Marina — three tenants, one surface.
Core capabilities
7 pillarsOpenAI-compatible gateway
A single `/v1/chat/completions` endpoint distributes traffic to any registered agent via `X-GoClaw-Agent-Id` + `X-GoClaw-User-Id`. Drop-in for any client that already speaks OpenAI.
Multi-provider fan-out
Anthropic, OpenAI, xAI, Gemini, Groq, OpenRouter, DeepSeek, MiniMax, DashScope, Claude CLI and OpenAI Codex — picked per agent, swapped without redeploy.
Subscription-backed inference
Mix pay-per-token APIs with flat-rate subscription plans — Claude Max, ChatGPT, z.ai — wrapped through Claude CLI and OpenAI Codex CLI adapters. The current production stack runs on a $200/mo Claude plan, a $100/mo ChatGPT plan, and adjacent subscriptions plus API top-ups — handling real workloads (dev tools, internal ops, batch jobs, WhatsApp concierge) that would invoice $2,000+/mo if every call hit token-metered APIs. The gateway treats subscription-backed providers as just another adapter — agents do not know which side the inference came from. Real-world API-grade use cases at a fraction of the per-token cost.
Agent registry in Postgres
Each agent — model, system prompt, context files, permissions — is a row. Hot-reload, no restart. New agent = INSERT.
3-tier memory + RAG
Working, episodic, semantic memory. pgvector + full-text hybrid retrieval over a per-agent knowledge vault. Cited, not hallucinated.
Tools + MCP grants
Think-act-observe loop with file, web, shell, and headless-browser tools. External tool servers attach over MCP (stdio / SSE / streamable-http) and are scoped per agent.
Multi-channel delivery
Same brain reachable from Telegram, Discord, Slack, WhatsApp, Feishu, Zalo and direct HTTP. One identity, many surfaces.
Powering, in production
jobXlaw
A curator agent maintains the bilingual corpus; a follow-up agent grounds turn 2+ on retrieved citations; a verifier agent runs stress tests against the labour-law index. All three share the same gateway and budget. Deployed on EC2, serving real labour-law queries.
Custom Cap BD (WhatsApp ops)
A private RedClaw deployment with full project-file access serves buyers and suppliers inside WhatsApp groups. Buyers ask for order status, fresh invoices, or PDFs on demand — the agent runs the generation scripts and posts the file back. Suppliers send shipment updates through the same channel; the agent ingests, updates the order record, and notifies the buyer. Same architecture as the public gateway, separate isolated instance for factory ops.
redxtrm /order
The `redxtrm-order` agent returns structured tool-calls that fill the wizard. The form stays source-of-truth; the gateway provides routing, auth, and per-tenant rate control.
Marina Rewards
`Rei Ai` answers Marina-only questions, with the same registry, auth surface and observability as everything else — but with a tenant-isolated context vault.
Image gallery
A sequenced walkthrough of the RedClaw control plane, ARC automation runs, Discord client operations, and tenant channels that share the same gateway.
Stack
ProductionNeed agents your team actually owns?
Self-hosted, single token surface, vendor-portable. Tell me what your agents need to do and which channels they need to live in — we’ll work backwards from there.