ara-sdk 0.1.19

Public Python SDK for building and running Ara apps.

Ara Python SDK

Public Python SDK for building Ara apps with a decorator-first workflow style.

Install

pip install ara-sdk

Local testing (no uv)

python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -e . pytest
python -m pytest -q

Principles

• Public SDK is generic and provider-agnostic.
• Runtime policy, retries, and safety controls are enforced server-side.
• Optional integrations (Cal.com, CRM, etc.) live in examples, not in the core SDK package.

Quickstart

from ara_sdk import App, Secret, invoke, runtime, schedule
import os

app = App(
    "Investor Meeting Booker",
    project_name="investor-meeting-booking",
    runtime_profile=runtime(
        secrets=[
            Secret.from_dotenv(),
            Secret.from_dict({"OPENAI_API_KEY": os.environ["OPENAI_API_KEY"]}),
        ],
    ),
)

@app.tool(id="send_email", description="Send one email.")
def send_email(to: str, subject: str, body: str) -> dict:
    return {"ok": True, "to": to, "subject": subject}

DAILY_FOLLOWUPS = schedule.cron(
    id="daily-followups",
    expr="0 13 * * 1-5",
    timezone="UTC",
    run=invoke.agent("booking-coordinator", input={"message": "Send pending follow-ups."}),
)

@app.agent(
    id="booking-coordinator",
    entrypoint=True,
    task="Coordinate scheduling requests.",
    skills=["send_email", "automation_create", "automation_list"],
    schedules=[DAILY_FOLLOWUPS],
)
def booking_coordinator():
    """Coordinate scheduling requests."""

export ARA_API_KEY="your_long_lived_api_key"
export OPENAI_API_KEY="your_provider_key"

ara deploy app.py
ara setup-auth app.py
ara run app.py --agent booking-coordinator --message "Need 3 slots next week"
ara run-async app.py --agent booking-coordinator --message "Need 3 slots next week" --response-mode poll
ara logs app.py
ara events app.py --event-type channel.web.inbound --channel web --message "hello"
ara setup app.py

If you prefer embedded script commands (python app.py deploy), add:

from ara_sdk import run_cli

if __name__ == "__main__":
    run_cli(app)

ara logs app.py streams live runtime events for the app across all active runs. Each line includes timestamp + run id + event type. To persist output, use shell piping:

ara logs app.py | tee app.logs

Environment

• ARA_API_KEY: long-lived user API key for control plane
  • In the Ara app, open Settings -> System, then use API Key -> Copy API Key.
  • Paste that value into ARA_API_KEY before running SDK commands.
  • Legacy ARA_ACCESS_TOKEN is still accepted as a compatibility fallback.
• ARA_API_BASE_URL: optional API override (defaults to production API)
• ARA_RUNTIME_KEY: optional runtime key override for run/events
• ARA_APP_HEADER_KEY: optional app header key override (X-Ara-App-Key) for run/events/run-async/run-status
  • Prefer running ara setup-auth app.py to mint keys and then export them.
  • Set ARA_APP_HEADER_KEY only when explicitly using X-Ara-App-Key mode.

Local bootstrap helper:

• ara setup-auth app.py:
  • resolves app_id by app slug
  • ensures a runtime key exists (optional)
  • creates /apps/{app_id}/x-keys key when missing
  • returns both runtime_key and app_header_key in command output

Runtime env and secrets

runtime(...) supports:

• env: plain runtime environment values (runtime_profile.env)
• secrets: ordered secret references (runtime_profile.secret_refs)

Container bootstrap (remote, Modal-style)

For container-side setup, declare startup bootstrap commands in runtime(...) and run them inside the provisioned app sandbox. This keeps build/bootstrap logic in the cloud container instead of depending on local uv workflows.

from ara_sdk import App, entrypoint, local_file, runtime

app = App(
    "Research Assistant",
    project_name="research-assistant",
    runtime_profile=runtime(
        image="python:3.12-slim",
        files=[
            local_file("./scripts/bootstrap.sh", path="scripts/bootstrap.sh", executable=True),
        ],
        startup=entrypoint("scripts/bootstrap.sh"),
    ),
)

python_packages / node_packages remain part of runtime profile metadata, but if you need deterministic install behavior today, use startup bootstrap commands in the container.

For observability, run warmup and stream runtime logs:

ara deploy app.py --warm true
ara logs app.py

Warmup/run lifecycle logs are emitted as run.warmup.* and run.*; startup failures are surfaced into those logs with command error previews.

Secret helper options:

• Secret.from_name(name, required_keys=None) (reference only)
• Secret.from_dict(name_or_env_dict, env_dict=None, *, required_keys=None, name=None):
  • Secret.from_dict("provider-local", {...}) for explicit naming
  • Secret.from_dict({...}) (or Secret.from_dict({...}, name="provider-local")) for programmatic local secrets
• Secret.from_dotenv(name=None, filename=".env") (auto-named when name omitted)
• Secret.from_local_environ(name, env_keys=[...]) (synced at deploy)

Deploy behavior:

• Local secret sources sync to /apps/{app_id}/secrets before warmup.
• When runtime(secrets=[...]) is present, deploy reconciles the remote app secret set to match those refs (stale secrets are removed).
• Secret references remain in manifest; plaintext values are not embedded in app manifest payloads.

Multi-sandbox proposal shape

The SDK can now declare sandbox placement and spawn intent in the manifest:

• policy: shared | dedicated | ephemeral | inherited
• key: logical sandbox selector used by runtime placement
• spawn: optional child-sandbox controls (to, max_depth, max_children_per_parent, max_total_child_sessions_per_run, ephemeral_ttl_minutes, child_policy, child_runtime)

Example:

sandbox(
    policy="dedicated",
    key="research-planner",
    allow_spawn=True,
    spawn_to=["deep-researcher", "verifier"],
    max_spawn_depth=3,
    max_children_per_parent=4,
    max_total_child_sessions_per_run=10,
    ephemeral_ttl_minutes=5,
    child_policy="ephemeral",
    child_runtime=runtime(memory_mb=1024),
)

Backward compatibility is preserved by default. Non-shared placement only activates when invocation input explicitly opts in:

• use_additional_sandbox=true, or
• sandbox.enable_additional_sandbox=true

Scheduling model

Use one schedule shape everywhere:

• schedule.cron(...) / schedule.every(...) for static declarations on @app.agent
• invoke.agent(...) / invoke.tool(...) for schedule targets
• scheduler.create(spec) for dynamic runtime automation payloads

Examples

See examples/ for optional integrations and demo projects:

• examples/calcom-booking/
• examples/async-ngrok-webhook/
• examples/programmatic-secrets-redeploy/ (live probe for secret reconciliation + stable generated secret names)
• examples/agent-skills-loading/ (three minimal Ara SDK skill execution apps)
• examples/framework-adapters/minimal_langgraph_subagent.py (legacy)
• examples/framework-adapters/minimal_agno_subagent.py (legacy)

Security

• Never commit API keys, runtime keys, or provider secrets.
• Keep provider-specific credentials in environment variables.

License

This repository is source-available under a strict proprietary license. Unauthorized copying, redistribution, or derivative works are prohibited. See LICENSE for full terms.