CrewAI decides who does what. AgentValet decides what they can reach.

A normal CrewAI tool holds the credential: to post to Slack, something in your process has a Slack token. With AgentValet the crew holds one identity key, and its tool list is built on the server from your grants, so a platform you haven’t granted produces no tool at all.

Livelast reviewed 24 September 2026

One pip install, one line in your crew

You need Python 3.10 or later. This path uses no MCP server and no Node at runtime.

1
Give the machine an agent identity. The register command prints four environment variables when it finishes. No Node? pip install agentvalet and run agentvalet register --code <code> with an invite or enrolment code; it writes the key to ~/.agentvalet/agent.key and prints the ids.
Terminal
```
npx @agentvalet/register
```
Set AGENT_ID and OWNER_ID where the crew runs, plus AGENT_PRIVATE_KEY_PATH if the key isn’t in ~/.agentvalet/.
3
Install the CrewAI package.
Terminal
```
pip install crewai-agentvalet
```

Hand your agent the governed tools. There’s one tool per platform you’ve granted, and nothing else.

crew.py

from crewai import Agent
from crewai_agentvalet import governed_tools

agent = Agent(
    role="Operations Engineer",
    goal="Keep the team informed",
    backstory="You handle routine ops chores end to end.",
    tools=governed_tools(),
)

Grant platforms in the dashboard. What you grant is exactly what the crew can see.

Use the MCP server instead

CrewAI speaks MCP too. Use the stdio server, not the remote URL: agent assertions last 60 seconds, and MCPServerHTTP sets its headers once, so a static token stops working a minute into the run. The stdio server signs a fresh assertion for every call.

Add the mcps= block to any agent in the crew.

crew.py

from crewai import Agent
from crewai.mcp import MCPServerStdio

agent = Agent(
    role="Operations Engineer",
    goal="Keep the team informed and the tracker up to date",
    backstory="You handle routine ops chores end to end.",
    mcps=[MCPServerStdio(
        command="npx",
        args=["-y", "@agentvalet/mcp-server"],
        env={
            "AGENT_ID": agent_id,
            "OWNER_ID": owner_id,
            "AGENT_PRIVATE_KEY_PATH": key_path,
        },
    )],
)

What’s governed here

Nothing is allowed until you grant it. This is a typical first grant for CrewAI; you decide every row, and you can change it without touching the agent.

Refusals come back to the agent as text it can reason about, not as exceptions that fail the task. When an approval is still pending, the tool tells the agent not to retry, because the action is already queued.

Example grant for CrewAI. Platforms and actions are illustrative.
What the agent tries	Platform	Decision
Read contacts	HubSpot	Allowed within grant
Post in #ops	Slack	Allowed within grant
Schedule a social post	Buffer	Asks you
Delete a contact	HubSpot	Not granted

Different roles, different authority

Give each role its own agent when their authority should differ. A researcher that only reads and a publisher that can post are two identities with two grants, and the audit log then tells you which role did what.

Your CrewAI logs say what the agents decided to do. The audit log says what they were allowed to do, and what a human said when asked.

CrewAI guide in the docsThe longer walkthrough, kept with the product docs.

All integrations or the platforms your agents can reach

Put CrewAI behind AgentValet.

Register the agent, grant only what it needs, and approve the risky calls from your phone. Free to start, no credit card.

Protect my first agent Watch your agent’s calls live, no signup