Jev in Hermes Agent

Hermes Agent is built to reason and act; Jev handles the hundreds of small, bounded decisions it makes along the way. Add Jev as an MCP tool and the agent gets a typed, calibrated choice / score / noul it can call mid-loop — a cheap second brain for the little picks, so the expensive model is spent on the work that actually needs it.

Hermes Agent (Nous Research) is a capable, open reasoning agent — but like every agent it burns a surprising amount of its loop on small structured decisions: which of these tools fits, is this action risky, which skill does this request need, does this result clear the bar to stop. Left to the main model, each of those is a full reasoning turn — slow, pricier, and returned as prose you still have to parse. Jev turns each one into a typed, calibrated call the agent can make in milliseconds and branch on directly, so Hermes spends its reasoning budget on the parts that need reasoning.

Add Jev as an MCP tool

The cleanest path is the open jev-mcp server. It runs straight from GitHub with npx — no clone, no build — and exposes Jev's decision primitives as MCP tools any MCP-speaking agent, Hermes included, can call. Point Hermes' MCP config at the command and set your key:

npx -y github:codaaiteam/jev-mcp
# set your key once — jev-mcp auto-routes jv_live_ keys to the hosted gateway:
export JEV_API_KEY=jv_live_...

Register that command as an MCP server in Hermes' tool config and restart. The Jev classify / score / check / gate tools appear in the agent's toolset — there's no endpoint or model to configure, because the jv_live_ key routes to the hosted gateway automatically.

The pattern: small decisions off the big model

The whole idea is a division of labour. Keep Hermes' main model for open-ended reasoning and generation, and hand the small, bounded, high-frequency decisions to Jev. A coding-style Hermes loop makes these constantly:

• Route: pick which tool or skill a request needs, out of your list (choice).
• Gate: allow / confirm / block a risky action before it runs (choice or noul).
• Score: rate quality, risk or priority on an ordered scale (score).
• Check: a calibrated yes/no with a probability behind it — 'is this done?', 'should I escalate?' (noul).

Each returns a typed value with per-option probabilities in ~70–500ms, so the agent acts on a number it can threshold instead of re-reading its own prose. And because the answer is calibrated — trained with RLCD to return honest probabilities — a rule like 'only auto-run if risk < 0.2' means what it says and stays stable between model versions, instead of tracking a number that drifts under you.

A concrete example

Say Hermes has 200 skills loaded and a request comes in. Instead of asking the main model to reason over all 200 (slow, and it may hallucinate a skill that doesn't exist), the agent asks Jev a single choice over the candidate skills and gets back one valid key with calibrated confidence. If confidence is low it falls back to the big model; if it's high it dispatches immediately. That's the shape of every good Jev-in-Hermes wiring: Jev takes the bounded pick, the expensive model takes the ambiguous 5%.

Why typed decisions beat a second reasoning turn

The alternative is to let Hermes' main model decide in free text. It works, but it's slow (seconds), pricier (dollars per million output tokens), and returns prose you have to parse — and it can invent an option that isn't in your list. Jev answers in 70–500ms at about $0.001 a decision, output free, and physically cannot emit an invalid type: a choice returns one of your keys and nothing else. Across the many small decisions inside an agent loop, that difference compounds into a faster, cheaper, more reliable agent.

Note

jev-mcp is a standard MCP server, so the same setup works in Claude Code, Cursor, OpenCode, Codex and any other MCP client — only the config format differs. See the Claude Code or OpenCode guides for those variants, or the agent-skill page for a client-agnostic install.

Preguntas frecuentes

Ver también: Jev MCP server · OpenCode integration · Jev for AI agents · Agent skill (all clients)