Install Eval Guardrails

Eval Guardrails is a paid guardrail service at eval.seanbehan.ca. Every installation puts a real hook in front of each tool call, so a dangerous action is evaluated before it executes, not just exposed as an MCP tool the model can ignore.

Harness	Interception seam	Install
OpenCode	plugin tool.execute.before	eval-jev install opencode
Hermes Agent	plugin pre_tool_call hook	eval-jev install hermes
DeepSeek Harness	bundle tools/pre-execute listener	eval-jev install dsh
Any MCP client	remote/stdio MCP tools	MCP-only fallback (no automatic interception)

Common steps

# 1. sign in and store the API key (or paste an existing eval_ key)
npx -y @codebam/eval-jev-guardrails login

# 2. install the hook for your harness
npx -y @codebam/eval-jev-guardrails install opencode
npx -y @codebam/eval-jev-guardrails install hermes
npx -y @codebam/eval-jev-guardrails install dsh --profile web

# 3. add credits when needed
npx -y @codebam/eval-jev-guardrails buy --pack p5000 --open

# 4. prove the hook is live
npx -y @codebam/eval-jev-guardrails doctor opencode

install prints the files and config it changed. Every command is idempotent; --uninstall reverses an install.

Decision mapping

Jev verdict	Hook behavior
allow	execute normally
review	deny by default (EVAL_REVIEW_MODE=allow to warn and continue)
block	deny, return the Jev reason to the model
support	deny, return the Jev reason to the model

One tool call costs one credit. Failed provider calls are refunded by the service. See pricing and the threat model for packs and limits.