Jev MCP Server: Connect the Jev Model to AI Agents

The AutoJev Jev MCP server makes structured Jev decisions available as tools that an AI agent can discover and call. It uses stateless Streamable HTTP at /mcp and accepts Bearer authentication.

What Jev MCP adds

Calling a model endpoint directly gives an application a decision response. A Jev MCP integration adds the operational pieces an agent needs:

• Tool discovery and typed input schemas.
• Stable names for common decisions.
• Portable Skills that define trigger conditions.
• A consistent Bearer authentication boundary.
• Preset guidance that maps decisions to workflow actions.
• A REST surface for clients that do not support MCP.

AutoJev MCP tools

The server currently exports six tools:

• autojev_route_model selects an allowed model using task stakes, quality, cost, latency, context, and tool-use requirements.
• autojev_guard_tool_call returns allow, confirm, review, or deny before a consequential tool call.
• autojev_route_task selects proceed_fast, deep_review, split_task, or block.
• autojev_check_research checks whether the supplied evidence supports a claim.
• autojev_review_completion checks whether work is complete and verified.
• autojev_decide evaluates custom Choice, Noul, and Score questions.

The tools return typed Jev answers and deterministic preset guidance. They do not execute the proposed action or expand the agent's authority.

Connect to the hosted Jev MCP server

An agent consuming AutoJev needs one AutoJev access key:

export AUTOJEV_API_KEY="your-autojev-access-key"

Add this Codex configuration:

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

Create a user-scoped key at https://autojev.ai/settings/apikeys. Keep it in a local credential store or the environment referenced by the client configuration. If a trusted local Agent reports a missing credential, you may explicitly ask it to configure AutoJev and supply the key in that private setup conversation. It should store the key locally without echoing it or putting it in a project file. Never send a key through a shared, public, or untrusted chat.

Install a Jev tool-call guard Skill

MCP tells an agent which tools exist. A Skill tells it when the decision boundary has been reached and how to handle the result.

mkdir -p .agents/skills/autojev-tool-guard
curl -fsSL https://autojev.ai/skills/autojev-tool-guard/SKILL.md \
  -o .agents/skills/autojev-tool-guard/SKILL.md

The Skill instructs the agent to call autojev_guard_tool_call before deployments, purchases, destructive changes, publishing, permission changes, and other difficult-to-reverse actions.

Security model

• Keep the AutoJev access key in an environment variable, not in a committed project file.
• Send only the task state needed for the decision.
• Never include passwords, access tokens, or unrelated customer data.
• Keep irreversible actions behind explicit human approval.
• Re-evaluate when the arguments, target, safeguards, or policy changes.
• Combine Jev judgments with deterministic validation and access control.

Use the online playground to inspect the exact request and response shape, or read the complete AutoJev integration guide.

Related Jev guides

• Start with What is Jev? for the Jev AI model overview.
• Read the Jev model guide for Choice, Noul and Score inputs and outputs.
• Install the umbrella router and five focused Jev Skills that trigger these MCP tools.
• See Jev AI for agents for routing, guardrail and verification patterns.