Many inputs.
One endpoint.
Zero black boxes.

Manifold sits between your tools and everything they call: it routes every model you run, turns documents and live databases into portable MCP servers, and governs it all with a single JWT — one subscription, zero infrastructure to run.

Open panel Documentation

docker compose up · one stack, seven services · nothing phones home

Request tracelive

$ curl your-host/v1/chat/completions -H "Authorization: Bearer $TOKEN"

route model=sirket-asistani → gpt-oss:120b node=gpu-01 healthy

retrieve project=support-kb chunks=4/8,412 hybrid 18ms

stream ttfb=214ms provider=ollama sse

200 1,284 tokens user=ali logged · audited

1endpoint

1JWT

0external calls

Platform

Three layers, one machine.

Model Gateway

Every provider behind one URL.

Ollama, vLLM, Bedrock and cloud gateways pooled with health checks, load balancing and two-layer failover. Format translation happens on the wire, so Cursor, Claude Code and Codex connect without plugins.

Knowledge Factory

Documents and databases, served as tools.

Upload documents or connect live databases; Manifold chunks, embeds and packages them as portable MCP servers — with hybrid retrieval and read-only SQL enforced at the query parser.

Orchestration

Soon

Tool order as infrastructure.

Compose MCP servers into a governed hierarchy where call order is enforced by the gateway itself — policy the model cannot talk its way around.

Integrations

Switching is a base URL.

The gateway answers in each tool's native dialect — OpenAI, Anthropic, Ollama and more — so your editors and agents keep their own SDKs. Copy the address from the panel's guide and you're done.

·Cursor
·Claude Code
·Codex
·OpenClaw
·Grafana
·VS Code

Point your tools here

# OpenAI-compatible (Cursor, SDKs)
OPENAI_BASE_URL=https://your-host/v1

# Claude Code & Desktop
ANTHROPIC_BASE_URL=https://your-host/claude

# Knowledge, as an MCP server
"mcpServers": { "support-kb": … }

Capabilities

Built like infrastructure, because it is.

Two-layer failover: A dying node or rate-limited model fails over to the next healthy option before the client notices.
One JWT across services: The same token authorizes the gateway, the knowledge servers and the panel — with per-user access and daily limits.
Read-only SQL, enforced: Live database access is parsed down to a single SELECT — writes are rejected before they reach your data.
Portable knowledge servers: Each project builds to a multi-arch OCI image with its vector store embedded. Run it anywhere Docker runs.
Live pipelines: Document processing and image builds stream their logs to the panel over SSE, token by token.
Private by default: Your models, documents and vectors live in an isolated workspace. Nothing is shared or exposed unless you choose.