Give a CrewAI crew secure secret access

The 1claw-crewai-tools package adds an OneclawVaultTool that fetches credentials from your vault at runtime using a scoped agent identity. Secrets are not copied into prompts, repos, or long-lived memory.

1. 1

   Install the tools package and CrewAI.

   bash Copy

   pip install 1claw-crewai-tools crewai

2. 2

   Export the agent identity, vault, and an LLM key. Use values from agent creation (see Agents & Access Control track) and your vault UUID from 1claw vault list.

   bash Copy

   export ONECLAW_AGENT_ID="paste-your-agent-uuid-here"
   export ONECLAW_AGENT_API_KEY="ocv_paste_your_agent_key_here"
   export ONECLAW_VAULT_ID="paste-your-vault-uuid-here"
   export GOOGLE_API_KEY="your-gemini-key"

3. 3

   Build a crew that gives one agent the vault tool. Save the script as crew_demo.py.

   python Copy

   import os
   from crewai import Agent, Crew, Process, Task
   from oneclaw_crewai import OneclawVaultTool
   
   vault_tool = OneclawVaultTool(
       agent_id=os.environ["ONECLAW_AGENT_ID"],
       api_key=os.environ["ONECLAW_AGENT_API_KEY"],
       vault_id=os.environ["ONECLAW_VAULT_ID"],
   )
   
   agent = Agent(
       role="Engineer",
       goal="Build features using vault-stored API keys",
       backstory="You use tools instead of pasted secrets.",
       tools=[vault_tool],
       verbose=False,
   )
   
   task = Task(
       description="Read path api-keys/openai using the vault tool; do not echo raw values.",
       expected_output="Confirmation that the path was read.",
       agent=agent,
   )
   
   crew = Crew(agents=[agent], tasks=[task], process=Process.sequential)
   crew.kickoff()

4. 4

   Create the secret if it doesn't already exist.

   bash Copy

   1claw secret set api-keys/openai --vault $ONECLAW_VAULT_ID --value 'your-openai-key' --type api_key

5. 5

   Run the crew.

   bash Copy

   python crew_demo.py

Your crew now fetches a secret from 1Claw through the oneclaw_vault tool during a task, without the value ever landing in code or repo.

What this gives you, and what it does not. Multi-agent frameworks multiply the surface, because every agent in the crew inherits whatever the crew's tools can reach.

• Tools are usually shared across the crew. An agent added later for an unrelated job often gets the whole toolset by default, including the vault access the first one needed.
• One agent's output becomes another's input, and it should be treated as untrusted. A researcher reading the web feeds text to a summariser, and that text can carry instructions.
• Cascading failure is the characteristic multi-agent bug. An early error is treated as established fact by every downstream step, which is why it appears in the OWASP agentic list as its own category.