Assistant OS from OpenClaw, decide layer from Jev. Routing, gates, and official decisionModel support.
What OpenClaw is
OpenClaw is an assistant operating system: multi-channel agents, tool execution, and pluggable providers. Third-party plugins extend tools and policies without replacing the host runtime. See https://docs.openclaw.ai for host documentation.
Official TypeSafe decision models in OpenClaw
0). The role stays disabled until selected. An optional typesafe_evaluate tool returns typed scores or probabilities after validation. Credentials use the host SecretRef pattern; billing follows TypeSafe API usage.
jev-claw: typed model routing
com) exposes jev_route: a single tool that classifies task_type, complexity, and risk, then returns route, confidence, needs_second_opinion, and reasons. Inputs include task text, changed_files, previous_attempts, and test_status.
Routing is a classification problem; Jev returns fixed labels with probabilities rather than free-form routing prose.
jev-harness: policy and gates
The jev-harness OpenClaw plugin wraps AntonioCoppe/jev-harness (npm library) for the decide layer: shell exec gates, tool allow/deny, model cost routing, inbox triage, browser next action, and stuck-agent recovery.
OpenClaw keeps generative work on Claude, Codex, or other providers; this plugin applies Choice/Score/Noul plus confidence thresholds and shadow mode.
{
plugins: {
entries: {
"jev-harness": {
enabled: true,
config: {
mode: "live",
gateExec: true,
gateTools: false,
toolAllowlist: ["web_search", "read"],
knownSafeCommands: ["ls", "pwd", "git status", "git diff"],
},
},
},
},
}Choose jev-claw vs jev-harness vs decisionModel
Use jev-claw when you need multi-model routing labels (cheap vs frontier, second opinion). Use jev-harness when you need pre-tool or pre-shell gates with recipes and shadow evals.
Use the bundled decisionModel provider when you want host-native typed evaluation via typesafe_evaluate without a separate plugin tool surface.
Environment keys
ai/ (or route through Vercel AI Gateway / OpenRouter where the integration documents it). Export TYPESAFE_API_KEY in the environment for OpenClaw plugin processes. Some MCP clients accept AI_GATEWAY_API_KEY when using gateway-hosted Jev models; follow that integration's README.
Never commit keys into repos or paste them into chat logs.
export TYPESAFE_API_KEY="tsk_..."
# Optional pin:
export TYPESAFE_MODEL="jev-latest"Pitfalls
0 and Node 24+. config/typesafe/api_key where the plugin supports it. jev-harness is not affiliated with TypeSafe or OpenClaw Foundation; read license and maintenance on GitHub. Start in shadow mode when tuning policies so Jev logs actions without changing live behavior.
Related directory listings
Related agent guides
FAQ
- Does Jev replace OpenClaw's chat model?
- No. Plugins and decisionModel roles handle structured decisions. Generative coding and conversation stay on your selected OpenClaw providers.
- Where is jev-harness source?
- The npm package and GitHub repo are github.com/AntonioCoppe/jev-harness. The OpenClaw plugin page describes how it maps recipes to plugin tools.
- What routes does jev_route return?
- openclawdir documents route labels such as cheap, main, architect, debugger, reviewer, claude-builder, claude-critic, and frontier, plus task_type and risk enums. Verify the current table in the plugin listing before hard-coding policies.
- Can I run both jev-claw and jev-harness?
- Yes, with clear ownership: routing vs gating. Avoid duplicate gates on the same action without coordinating policy order in OpenClaw config.