Drop in Dream-Weaver and route across local GPU inference, Anthropic, OpenAI, Gemini, and a dozen others — with a single OpenAI-compatible endpoint. Your data never leaves unless you choose.
https://api.dream-weaver.ai/v1
Full OpenAI API compatibility — chat completions, embeddings, streaming, function calling. Works with any SDK that speaks OpenAI.
A mixture-of-experts classifier reads each request and routes: code to Codestral, reasoning to qwen3, long-form to Claude — automatically. Or override per-request. Cost drops, quality rises.
Semantic memory with pgvector and 1024-dim mxbai embeddings. Your codebase, your notes, your conversations — searchable across every model and every session, stored on your hardware.
Route sensitive workloads to your local GPU (NVIDIA or AMD). Route cost-insensitive tasks to cloud. Your data doesn't move unless the routing policy says so — you set the policy.
The dream-weaver/auto model ID runs a lightweight classifier on every incoming request to determine the optimal backend: local GPU, Anthropic, OpenAI, or Ollama. Routing decisions are logged and auditable. Cost reduction with no accuracy trade-off for well-classified workloads.
Embeddings generated by mxbai-embed-large (1024 dims, pgvector) surface relevant memory at inference time. Every conversation, document, and code snippet stored on your hardware — not in a vendor's cloud. Searchable via the /v1/memories API or MCP tools.
An optional reasoning pod (LSR) sits between the request and the model. For structured queries — rule-following, constraint satisfaction, fact checking — LSR augments the LLM with Prolog-style symbolic rules via inductive logic programming. Working memory persists across calls.
Dream-Weaver runs inside your Kubernetes cluster. The routing gateway is self-hosted. Memory lives in your Postgres. Sensitive requests never leave your network unless you explicitly route them to a cloud backend.
Ollama backend supports CUDA (RTX, A100, H100, GB200) and ROCm (MI300X, MI355X). Bring your own iron, run your own models.
Define per-user, per-team, per-model-id routing policies. Compliance-sensitive traffic stays local. Everything else can use the cheapest capable model.
Full request/response audit log in Postgres. See what model answered, what it cost, how long it took. No black box vendor dashboards.
Keep your same OpenAI SDK calls. Route cheap workloads to local qwen3 or Codestral. Reserve Claude Opus for the tasks that actually need it. Watch your monthly bill drop.
Healthcare, finance, legal — where data residency isn't optional. Deploy Dream-Weaver in your VPC. Your teams use the same developer experience, your InfoSec keeps their guarantees.
If you're running MI300X or H100 clusters, your idle inference capacity is valuable. Dream-Weaver turns your HPC investment into an internal LLM API — for your whole organization.
Persistent semantic memory, working memory via LSR, and MCP-compatible tools mean your agents remember across sessions without you building the plumbing. Focus on the agent logic.
Dream-Weaver is currently available to early-access teams. If you're running local GPU hardware or want OpenAI-compatible access to every frontier model — get in touch.