Kronaxis Platform
One infrastructure layer, two independent products. Kronaxis Router is the cost-routing OpenAI-compatible proxy with built-in RAG and Claude Code wrapping. Kronaxis Fabric is the shared semantic-memory + cross-session coordination MCP on Postgres. Install one. Install both. Each stays independently usable; together they compose into a single self-hosted infrastructure layer for any multi-agent LLM workflow. (The capability-typed task queue arrives in Fabric v0.6 and slots in without rewiring.)
The dotted line between Router and Fabric is the composable bit. Set fabric_url: in the Router config and the RAG pre-stage stops using its embedded pgvector and starts asking Fabric instead. One memory store, two consumers.
Install one. Install both. Same upgrade path.
Like HashiCorp ships Vault, Consul, and Nomad as independent products that compose under one operational story, Kronaxis ships Router and Fabric as independent binaries that compose into one infrastructure layer. Pick the deployment shape that matches today, not the one you might need next year.
A cost-routed LLM proxy with an embedded pgvector RAG pre-stage and Claude Code wrapped behind /v1/chat/completions. No Fabric needed: bring your own Postgres, point Router at it, ship. This is the default behaviour and the smallest box of moving parts.
A single Go binary that turns one Postgres into shared semantic memory, full-text+vector hybrid search, cross-session coordination over pg_notify, and an MCP server your agent fleet talks to directly. No LLM proxy, no cost router. Just the memory backplane.
Set fabric_url: in Router's config and the RAG pre-stage starts pulling chunks from Fabric instead of embedded pgvector. Now agents (via MCP) and Router (via /v1/rag) read from the same memo store. Bank a memo from Claude Code — your next API call through Router can cite it. One upgrade path.
How Kronaxis Platform compares
Three honest comparisons. If your problem is exactly one of these stacks' sweet spots, use that stack. If your problem is "I need an LLM proxy and a memory backplane on the same Postgres, self-hosted, with the option to add a task queue later", that's what we ship.
| Kronaxis Platform | HashiCorp suite (Vault + Consul + Nomad) | LangChain stack | LlamaIndex + LiteLLM | |
|---|---|---|---|---|
| Independently installable products | ✓ | ✓ | partial (one Python package, many modules) | ✓ (two projects) |
| Self-hosted, single-binary per product | ✓ (Go) | ✓ (Go) | Python service | Python service |
| LLM cost routing across providers | ✓ (Router) | × | partial (manual) | ✓ (LiteLLM) |
| Shared semantic memory + cross-session coord | ✓ (Fabric) | service discovery only | memory primitives, not a service | retrieval store only |
| One Postgres backs the whole platform | ✓ | × (each has its own store) | × | × |
| Claude Code CLI as an OpenAI endpoint | ✓ | × | × | × |
| Hybrid rank (cosine + tsvector + recency) default | ✓ | × | configurable, not default | configurable, not default |
| MCP-native tool surface for agents | ✓ (Fabric) | × | × | × |
| Capability-typed task queue | ✓ (Fabric v0.6) | ✓ (Nomad) | × | × |
| Best for | self-hosted multi-agent LLM ops | general infra orchestration | building agent apps in Python | RAG over your docs at scale |
Two products, two repos, one platform
OpenAI-compatible proxy with cost routing, pgvector RAG, multi-account auth pool, and Claude Code wrapping. 9.9 MB Go binary, 22K req/s, 5ms p50.
MCP-native memory + coord service on Postgres. 22 endpoints, hybrid retrieval, pg_notify coord, tree-sitter code graph. v0.6 adds the capability-typed task queue.