Your localhost, on one screen.

Every tool in the catalog, plotted by the port it wants.

The dashboard is the catalog. The gateway is the part that does things.

It holds the Docker socket, reads your hardware, discovers what Ollama has loaded, and exposes one REST surface that the UI consumes. Here is that surface, endpoint by endpoint — every path below exists in gateway/routes/.

• Containers

  dockerode, never a shell

  • GET /api/services
  • POST /api/services/:name/start
  • POST /api/services/:name/stop
  • POST /api/services/:name/restart
  • GET /api/services/:name/logs

  Start, stop, restart and tail any container from the dashboard. Names are validated against /^[a-zA-Z0-9][a-zA-Z0-9_.-]*$/ before they reach the daemon; a container name has never touched a shell since v2.0.0.

• Hardware telemetry

  polled and streamed

  • GET /api/telemetry
  • GET /api/telemetry/stream

  CPU utilisation, cores and model; system RAM; disk; the Node process’ own memory; GPU and VRAM utilisation and temperature via nvidia-smi, with a fallback when there is no NVIDIA card.

• Models

  Ollama and vLLM, discovered

  • GET /api/models
  • GET /api/models/active
  • POST /api/models/switch
  • POST /api/models/preload

  Finds what is loaded on :11434 and vLLM, reads parameter sizes, quantisations and families, and keeps a model warm with a configurable keep_alive so the next switch has no cold start.

• OpenAI-compatible layer

  point your existing client at :3001

  • GET /v1/models
  • POST /v1/chat/completions
  • POST /v1/completions
  • POST /v1/embeddings

  Chat completions stream over SSE or return whole. Anything that already speaks the OpenAI wire format can be aimed at the gateway instead of at api.openai.com.

• RAG, offline

  no embedding API involved

  • POST /api/rag/index
  • POST /api/rag/query
  • POST /api/rag/search
  • GET /api/rag/documents/:id/chunks

  Sliding-window chunking on sentence boundaries, 384-dimensional L2-normalised vectors generated locally (with an Ollama embedding fallback), cosine similarity search, and answers synthesised with source citations.

• Benchmarks

  numbers you produce, not numbers we print

  • POST /api/benchmarks/run
  • GET /api/benchmarks
  • GET /api/benchmarks/:id

  Measures time-to-first-token, tokens per second, total latency and memory footprint across one model or several, comparatively. The results are yours; this page does not quote any.

• Agents & workflows

  execution, not CRUD theatre

  • POST /api/agents/:id/task
  • GET /api/agents/:id/tasks
  • POST /api/workflows/:id/execute
  • GET /api/workflows/:id/runs

  Agent tasks run against your local Ollama. Workflows run prompt, http and delay steps in sequence, a failed step stops the run, and history is capped at 20 tasks and 10 runs per record.

• Event stream

  server-sent, with heartbeats

  • GET /api/events

  A Docker snapshot every five seconds plus task and workflow completion events, so a client can stop polling.

It also tells you when it cannot tell. Health probes run from the browser, so a service with no CORS headers reports CORS blocked rather than pretending to be down, and a TCP-only check reports unknown rather than guessing. That distinction is in the code, in the tests, and on the badge.

Eight layers is not a marketing count. It is the index.

Every tool in the catalog belongs to exactly one layer, and the dashboard's pages, filters and per-layer counts are all derived from that single field. Change the field, the whole app re-sorts itself. Scroll, and the plan view pulls apart.

Forty-one tools, and the port each one answers on.

Ports, docs links, start hints and health-check config for every entry live in one TypeScript file. Filter by layer:

Clone it, install it, open :3000.

# 1 — get it
$ git clone https://github.com/Arnav1771/localhost-ai-hub.git
$ cd localhost-ai-hub

# 2 — node check, .env files, deps, data dirs
$ ./scripts/init.sh

# 3 — two terminals
$ npm run dev:gateway     # :3001
$ npm run dev:dashboard   # :3000

$ docker compose up --build -d
$ curl localhost:3001/health

The unflattering half of this page.

There is a v3 specification for this project that describes a graph runtime, artifacts, evaluations and human-in-the-loop approvals. Almost none of it exists yet. Here is the split, so nothing on this page has to be taken on faith.

A desktop build — one executable instead of two npm processes — is the direction, not a release. It has not been started. Today it is npm run dev:gateway and npm run dev:dashboard, or a two-service compose file, and this page will say so until that changes.