Question 1

How is this different from just calling OpenRouter or Ollama directly?

Accepted Answer

It is a self-hosted façade: one base URL, one compatible request shape, and provider selection per request instead of per client. That earns its keep when you are attaching it to something like Bolt.new, v0 or Lovable that cannot hold several keys or choose a provider. It adds no model quality. For anything past text, vision and data endpoints you are routing to the same upstream models you could already reach yourself.

Question 2

Does the agent actually call tools, or is it a chat wrapper with extra steps?

Accepted Answer

It calls tools. The agent plans across turns, picks tools, reads the results and keeps going. Scenarios group a toolset — general, customer-support and real-estate — and ship with demo data so the demos work immediately. Session memory is why a follow-up like asking whether the last payment went through resolves. Both endpoints stream over SSE. You can also register custom HTTP tools from the dashboard with a JSON Schema for arguments, and MCP servers over Streamable HTTP or SSE.

Question 3

What does Jev do here that a normal text provider cannot?

Accepted Answer

Jev is listed in a separate section as an evaluation and decision provider, explicitly not a text generator. You send a piece of state plus typed questions — choice, score or noul — and get a selected option with calibrated probabilities, which suits routing, triage and thresholds rather than prose. It needs its own key, base URL and model variables and a dedicated evaluate endpoint. No latency, cost or accuracy figures are given for it.

Question 4

What are the security catches before I self-host this?

Accepted Answer

The admin surface is the main one: it can register MCP servers and arbitrary HTTP tools, and it writes provider keys to a gitignored JSON file on disk. Response masking covers the keys but the file does not, so treat it as a secret store. All admin APIs require a bearer token in production. The bundled agent scenarios execute against demo data only. Access token and CORS variables exist, but the README says nothing about rate limiting, audit logging or sandboxing of tool calls.

Question 5

Which providers does it front?

Accepted Answer

Local Ollama, LM Studio, and LlamaCpp, plus LLM Gateway, OpenAI, Anthropic, OpenRouter, Vercel AI Gateway, Google AI Studio, Baseten, and ZAI, all behind one base URL and request shape.

Question 6

What is the admin dashboard for?

Accepted Answer

It runs at /admin and lets you create and edit agents, define custom HTTP tools with a JSON Schema for arguments, register MCP servers whose tools are auto-discovered, create skills, and set provider API keys at runtime.

Question 7

How long do agent sessions last?

Accepted Answer

The chat endpoint keeps multi-turn sessions in memory with a 30-minute idle expiry, so anything durable needs your own storage; the one-off endpoint returns the final answer plus every tool call it made.

AI Backends

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community