pushary-crewai

Phone approvals for CrewAI agents. Your agent asks, your user taps Approve or Deny.

Full walkthrough: Human-in-the-loop for CrewAI

What you need

• A Pushary Partner plan, from $99 a month. Start the trial.
• An API key from Partner onboarding, set as PUSHARY_API_KEY.
• Your users install the free Pushary app (iPhone, Android). They never sign up or pay.

Quick start

pip install pushary-crewai
export PUSHARY_API_KEY=pk_xxx.sk_xxx

from crewai import Agent, Task, Crew
from pushary_crewai import connect, make_ask_human_tool

link = connect("user_123")  # once per user: show them this link

agent = Agent(
    role="Ops",
    goal="Ship safely",
    backstory="Careful operator.",
    tools=[make_ask_human_tool("user_123")],
)
task = Task(
    description="Draft the release. Before finalizing, call ask_human to get approval.",
    expected_output="approved release notes",
    agent=agent,
)
Crew(agents=[agent], tasks=[task]).kickoff()

This replaces the console human_input=True prompt. The tool waits until the person answers on their phone. A no, or no answer, tells the agent not to proceed. The external_id is bound when you build the tool, never taken from the model, so a prompt-injected agent cannot ask the wrong person.

CrewAI's own docs on this: Human feedback in Flows.

Lower-level helpers

from pushary_crewai import ask_human

d = ask_human("Approve this refund?", external_id="user_123", type="confirm")
if d["approved"]:
    issue_refund()

API

• connect(external_id, *, api_key=None, base_url=None): enroll an end-user's phone.
• make_ask_human_tool(external_id, *, name=..., ...): a CrewAI BaseTool bound to that user.
• ask_human(question, *, external_id, type="confirm", ...): blocking, returns the decision dict.
• resolve_pushary_callback(raw_body, signature, secret): verify + parse a callback for a durable path.
• describe_answer(type, result), is_affirmative(answer), deterministic_key(parts), SIGNATURE_HEADER.

Example

The basic example gives an agent a blocking ask-human tool.

The Flow example uses CrewAI's native @human_feedback(provider=...) path. It creates a Pushary decision, saves the pending Flow, and exits. A later command reads the decision from Pushary and resumes the same Flow. No model key or open worker is needed. Tested with crewai==1.15.22 and pushary==2.1.1.

python -m pip install -e . 'crewai==1.15.22'
export PUSHARY_API_KEY='your Partner server key'
export PUSHARY_EXTERNAL_ID='your enrolled customer ID'
export CREWAI_STORAGE_DIR="$(mktemp -d)"
python examples/flow_feedback.py start 'Send the release note'
# Answer on the connected phone. Copy flow_id and decision_id from the output.
python examples/flow_feedback.py resume FLOW_ID DECISION_ID \
  --database "$CREWAI_STORAGE_DIR/resumes.sqlite"

Keep CREWAI_STORAGE_DIR and the resume database across processes. A pending decision returns pending; a decline or expiry returns approved: false. The example checks the decision ID, exact question, Flow, and recipient before resuming. Its SQLite claim prevents two local workers from resuming the same Flow concurrently; a failed or interrupted resume stays claimed for manual inspection. Use shared transactional storage across hosts.

The Flow only returns the review result. If you attach a business action, bind its exact arguments and recipient in your application, and make the action idempotent. Do not treat a model's approval claim or a direct call to flow.resume() as authorization.

License

MIT