Jev Skills for AI Agents: Install and Use AutoJev Skills

Install the AutoJev tool router and focused Jev Skills for model routing, task routing, tool-call protection, research verification and completion review.

Last updated: 2026-09-19

Jev Skills are portable instruction files that teach an AI agent when to ask the Jev model for a decision, what task state to send and how to handle the typed result. They work alongside the AutoJev Jev MCP server, which exposes the callable tools.

MCP answers “what tools are available?” A Skill answers “when should the agent use this tool, and what should it do with the result?”

The six AutoJev Skills

AutoJev Tool Router

autojev maps a decision boundary to one of all six AutoJev MCP tools. Install it when you want broad AutoJev coverage without loading every focused workflow Skill. It prefers the five preset tools and uses autojev_decide only for a custom typed decision.

Jev Task Router

autojev-task-router helps an agent choose whether to proceed quickly, review deeply, split a task or block. It is intended for work with meaningful ambiguity, risk, weak verification or irreversible consequences.

Jev Model Router

autojev-model-router selects among models the current environment can actually invoke. The Skill asks the agent to describe task stakes, quality, cost, latency, context and tool-use requirements without inventing benchmarks.

Jev Tool Guard

autojev-tool-guard evaluates a proposed consequential tool call immediately before execution. It covers external mutations, spending, data exposure, publishing, permission changes, deployment and other difficult-to-reverse actions.

Jev Research Guard

autojev-research-guard checks whether collected evidence supports one precise claim. It is useful for conflicting sources, weak provenance, current facts and high-stakes conclusions.

Jev Completion Review

autojev-completion-review compares an objective with completed work, verification, failures and known gaps before an agent reports success.

Install all AutoJev Skills at once

Run this once from the project root, or ask your local coding agent to run it. The loop installs the umbrella router and all five focused Skills into the project's .agents/skills directory:

bash

for skill in autojev autojev-task-router autojev-model-router \
  autojev-tool-guard autojev-research-guard autojev-completion-review; do
  mkdir -p ".agents/skills/$skill"
  curl -fsSL "https://autojev.ai/skills/$skill/SKILL.md" \
    -o ".agents/skills/$skill/SKILL.md"
done

After installation, ask the agent to read .agents/skills/autojev/SKILL.md and configure AutoJev for the project. Review Skills before installing them into a trusted user or project directory. The files are readable instructions, so a team can version, audit and customize them like other project policy.

Connect the Jev MCP server

The Skills expect the corresponding AutoJev MCP tools to be available. For Codex, put the AutoJev access key in the environment and add:

toml

[mcp_servers.autojev]
url = "https://autojev.ai/mcp"
bearer_token_env_var = "AUTOJEV_API_KEY"
tool_timeout_sec = 30

Keep the AutoJev access key in an environment variable rather than a committed MCP configuration or Skill file.

See the Jev MCP guide for additional clients and tool details.

How Jev Skills trigger decisions

A useful Jev Skill contains more than a tool name. It defines:

Trigger conditions that identify a real decision boundary.
Context preparation that keeps state compact and excludes unrelated data.
Tool selection that maps the situation to one typed Jev MCP tool.
Verdict handling that explains what to do with allow, confirm, review, deny or other preset results.
Safety invariants that preserve user authority, permissions and required approvals.

For example, the tool-guard Skill triggers immediately before a consequential call, not at the beginning of an unrelated planning session. It sends the proposed action, side effects, safeguards and policy, then applies the returned verdict without treating it as new authorization.

Jev Skills do not expand authority

Installing a Skill does not grant file access, shell access, deployment rights or permission to spend money. The agent remains limited by its host environment, the user's instructions and existing approval rules.

Jev Skills should never be used to:

Bypass a confirmation already required by the user or platform.
Send passwords, API keys or unrelated private data to a decision tool.
Treat model confidence as proof of correctness.
Replace deterministic validation that can be performed reliably in code.
Execute an irreversible action solely because a model returned allow.

Which Jev Skills should I install?

Workflow need	Recommended Skill
Route across all six AutoJev tools	`autojev`
Choose an execution path for ambiguous work	`autojev-task-router`
Select among available AI models	`autojev-model-router`
Check a consequential tool call	`autojev-tool-guard`
Validate whether evidence supports a claim	`autojev-research-guard`
Check that a task is complete and verified	`autojev-completion-review`

The command above installs all six Skills and is the recommended default. For a minimal host, you can instead download only one row from the table into .agents/skills/<skill-name>/SKILL.md. Focused Skills use narrower triggers; the umbrella autojev Skill provides broad tool discovery.