stillrunning-sdk

Monitoring for any Python job, agent, or script, in one line.

Decorate or wrap your code and StillRunning gets a ping on success or failure with the run's duration, plus optional cost, tokens, and model. Get alerted the moment a cron job stops, an agent fails, or a script runs too long or costs too much, with no ping plumbing. Works with LangChain, CrewAI, agent scripts, Celery/RQ workers, and plain cron jobs.

pip install stillrunning-sdk

30-second quickstart

1. Create a workflow at stillrunning.ai/app/new and copy its token.

2. export STILLRUNNING_TOKEN=your_token_here

3. Monitor your work:

   from stillrunning_sdk import StillRunning
   
   sr = StillRunning()  # reads STILLRUNNING_TOKEN
   
   @sr.monitor
   def nightly_job():
       ...
   

Every call to nightly_job is now timed and reported, success on return, failure on exception (re-raised untouched).

Three ways to use it

Decorator , monitor a function:

@sr.monitor
def sync_customers():
    ...

Context manager , attach AI metrics inside the block:

with sr.track() as run:
    result = call_agent(prompt)
    run.meta(tokens_in=result.usage.input, tokens_out=result.usage.output, model="gpt-4o")
    # model + tokens -> costUsd estimated automatically

Run a callable , inline:

answer = sr.run(lambda: call_agent(prompt), meta={"model": "claude-sonnet-4", "tokens_in": 1200})

Heartbeats and manual pings

For a cron job that just needs to say "I ran", or when you can't wrap the work:

sr.heartbeat()                          # bare success ping
sr.ping(event="start", traceId="...")   # low-level: success | fail | start | log

Cost estimation

Give model plus token counts and cost is estimated from a built-in pricing table (Claude / GPT / Gemini). It's approximate, it powers relative cost-anomaly detection and a ballpark spend figure. Pass an explicit cost_usd for exact accounting, or extend the table:

from stillrunning_sdk import register_model_pricing
register_model_pricing([(r"my-custom-model", (1.5, 6.0))])  # USD per 1M (input, output) tokens

Unknown models simply send no cost rather than a wrong one.

Grouping multi-step runs with with_trace

from stillrunning_sdk import StillRunning, with_trace
sr = StillRunning()

with with_trace():
    sr.run(plan_step)
    sr.run(execute_step)   # both pings share one trace_id

Configuration

StillRunning(
    token=None,        # defaults to env STILLRUNNING_TOKEN
    base_url="https://stillrunning.ai",
    timeout=3.0,       # seconds; bounds each ping so a slow StillRunning never hangs your job
    await_ping=True,   # False = fire-and-forget in a daemon thread (lowest latency)
    on_error=None,     # callable(Exception) -> observe ping delivery failures
    transport=None,    # custom transport (testing / non-urllib runtimes)
)

Monitoring never raises into your code: a failed ping routes to on_error and is otherwise swallowed.

Requirements

Python 3.9+. No runtime dependencies (uses the standard library).

License

MIT