Four jobs, four kinds of integration
| Job | What Jev decides | What still belongs to Hermes |
|---|---|---|
| Skill selection | Which skill or skills are worth loading for the current task. | Loading them and executing anything they enable. |
| Command approval | Whether a pending shell or tool request looks risky. | The permission boundary and the final allow or deny. |
| Model or tool routing | Which model, handler or tool should take the step. | Making the call and handling its output. |
| Context control | Which tool results are worth keeping when context is trimmed. | The transcript Hermes finally sends to the model. |
This is the Hermes-specific detail behind the coding-agent integrations overview. Jev judges; Hermes owns the environment.
Skill selection and routing
Hermes Jev Skills is the broadest entry point: its README documents model routing, memory and compaction decisions, skill selection and computer or browser use. Enable the component that addresses your bottleneck instead of turning everything on at once.
Two narrower projects focus only on skill choice. ask-jev-skill documents a skill that lets Hermes ask Jev a question, and typesafe-skill-router documents naming the one skill worth loading before the model call. Your code still owns the available catalog and the "no suggestion" outcome. The official skill-suggestion recipe shows the same shortlist-and-check pattern.
Command approvals and risk review
Approval is the highest-value Hermes job for a decision model, and also the one where wording matters most. hermes-jev-approvals documents using Jev as a reviewer for smart command approvals. jev-decisions-hermes documents tool-risk reviews with human-approval recommendations, and hermes-jev-plugin exposes decision tools such as check, route, score and evaluate.
None of these make a model the access-control boundary. Treat the output as a review signal that feeds a deterministic rule, and keep an explicit path for a human to decide. The guardrails guide covers that boundary in detail.
Model and tool routing
hermes-jev-router documents compacting tool results, suppressing duplicate tools and skipping unnecessary steps. jev-router documents decision routing plus a session-start skill pre-loader, using OpenRouter as the transport. Both are community plugins: read whether each one adds a gate, preserves an existing gate, or disables itself when a key is missing.
A routing decision is not a code-generation decision. Selecting a cheaper or more capable model does not turn Jev into the model that writes the code, which is the first distinction in the overview.
Context and compaction
jev-compaction documents scoring every tool call instead of summarising Hermes sessions. That is a retention decision: which spans survive. Compare any project against the context compaction guide and test whether the information a task needs actually survives the trim.
What to check before installing
- Confirm which Hermes release and plugin API the project targets; hosts change.
- Decide whether the plugin needs a TypeSafe key, an OpenRouter key or neither.
- Check what happens on a timeout, a malformed answer or a missing key; fail-open and fail-closed differ.
- Run in an observation-only mode first, where the project supports one, and compare its proposals with what the task needed.
Failure modes and boundaries
Model output is not a security control, and "auto mode" is not one behavior. Read each README for how a plugin behaves without credentials, during an outage and on ambiguous input. Start with a narrow job such as skill selection, log the decisions, and only then widen the plugin's authority. If you mainly need routing or tool approval rather than a Hermes-specific integration, the same decision layer appears for Pi in the Pi integration guide.