Question 1

Do I have to run a server somewhere, and what does it expose?

Accepted Answer

Yes. The documented architecture has the client POST to /mcp on a daemon reachable over Tailscale, authenticating with a bearer token, and that daemon calls the System One endpoint over HTTPS with the TypeSafe key. It exposes six tools — classify, check, rate, batch_judge, system_one_raw and health — plus server-side confidence gating, cost caps, caching and a decision log. Same-machine development runs over stdio instead.

Question 2

Why is check() reporting a confidence the API never sent?

Accepted Answer

Because noul answers come back with no confidence field, and the server derives one locally as the distance from 0.5 doubled. The README records this as a known limitation and notes the effect is that noul reads as more conservative than choice by design. If you would rather have the raw response, system_one_raw is the passthrough escape hatch, though the model id stays pinned there too.

Question 3

What happens to my data, and what gets logged?

Accepted Answer

The TypeSafe key is read from ~/.jev/.env or the environment, and the client token from ~/.jev/token, both at mode 0600 with the token generated by openssl rand. Every decision is appended to a JSONL file under the local share directory, also at 0600 and metadata-only by default; setting JEV_LOG_FULL_RESPONSE=1 opts into writing response bodies. Both secrets are gitignored, and HTTP mode fails fast at startup if the client token is missing.

Question 4

Will other deployments hit the same walls this one did?

Accepted Answer

The README keeps a table of pitfalls it fixed, each with a regression test: FastMCP's transport_security rejecting a Tailscale host with a 421, the VSCode extension ignoring mcpServers in settings.json, and httpx trust_env=True causing intermittent TLS failures behind a local proxy with an empty-looking error string. It also records cutting the batch cap from 50 to 20 after fifty 300 ms calls tripped the MCP client timeout, and that httpx is pinned to trust_env=False.

Question 5

What tools does the server expose, and what are their limits?

Accepted Answer

classify (2 to 255 options), check (yes/no with a server-computed effective confidence), rate (2 to 10 ordered levels), batch_judge (independent items capped at 20), system_one_raw as a passthrough with the model pinned to jev-1.13.0, and health, which runs a real GET /v1/models probe plus session stats.

Question 6

How are costs bounded per session?

Accepted Answer

A 32,000-character state cap and a $0.10 per-session budget, set by JEV_STATE_CAP_CHARS and JEV_SESSION_BUDGET_USD. There is also a sha256 LRU cache of 1000 entries with no TTL, and model retries only cover 429 and 529 with exponential backoff.

Question 7

How is the server deployed and secured?

Accepted Answer

As a systemd user service jev-mcp.service, enabled with Linger=yes, binding the Tailscale IP on port 8765 only rather than 0.0.0.0, with bearer auth via JEV_MCP_TOKEN. HTTP mode fails fast at startup if that token is missing, so an unauthenticated server never starts.

Jev MCP — Songokou1983

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community