stillrunning-claude-agent-sdk

One-line monitoring for the Claude Agent SDK. Wrap query() and every agent run auto-reports to StillRunning with its duration, tokens, cost, model, and tool-call count, and fires an alert the moment a run fails, runs too long, or costs too much.

import { query } from '@anthropic-ai/claude-agent-sdk'
import { stillrunning } from 'stillrunning-claude-agent-sdk'

const sr = stillrunning() // reads STILLRUNNING_TOKEN
const monitoredQuery = sr.wrap(query)

for await (const message of monitoredQuery({ prompt: 'fix the failing test' })) {
  // use messages exactly as before , a ping fires automatically when the run completes
}

That's the whole integration. monitoredQuery is a drop-in for query: same arguments, same message stream, same control methods (interrupt(), setPermissionMode(), …). It just watches the run go by.

Install

npm install stillrunning-claude-agent-sdk

The Claude Agent SDK is your dependency; this package wraps the query you already import, so there's nothing to keep in version lockstep.

Setup

1. Create a workflow at stillrunning.ai/app/new and copy its ping token.
2. Set STILLRUNNING_TOKEN (or pass stillrunning({ token })).
3. Wrap query once and use it everywhere.

What it captures

On completion the wrapper reads the Claude Agent SDK's final result message and pings StillRunning with:

| Field | Source | | --- | --- | | success / fail | result.subtype (error_* or is_error → fail) | | duration | result.duration_ms | | cost | result.total_cost_usd (the SDK's own number) | | tokens in / out | result.usage (input + cache reads/writes, and output) | | model | the dominant model in result.modelUsage | | tool calls | counted from the tool_use blocks in the run | | turns, session id, terminal reason | sent as metadata |

A run that throws mid-stream pings fail with the error message, then the error rethrows untouched. A failed run still reports the tokens and cost it burned before failing.

Grouping multi-step agents

Wrap several runs in withTrace so they share one traceId and stitch into a single outcome chain in StillRunning:

import { stillrunning, withTrace } from 'stillrunning-claude-agent-sdk'
const sr = stillrunning()
const query = sr.wrap(baseQuery)

await withTrace(async () => {
  for await (const _ of query({ prompt: 'plan the work' })) {}
  for await (const _ of query({ prompt: 'execute step 1' })) {}
}) // both runs share one traceId

Configuration

stillrunning({
  token,         // ping token; defaults to process.env.STILLRUNNING_TOKEN
  baseUrl,       // defaults to https://stillrunning.ai
  computeCost,   // ({ model, inputTokens, outputTokens }) => number , override the SDK's cost
  awaitPing,     // default true; false = fire-and-forget (lowest latency)
  pingTimeoutMs, // default 3000
  onError,       // (err) => void , observe ping delivery failures
  fetch,         // custom fetch (testing / non-global-fetch runtimes)
})

Monitoring never throws into your code: a failed ping routes to onError and is otherwise swallowed, bounded by pingTimeoutMs so a slow StillRunning never hangs your agent. Messages stream through unchanged.

Requirements

Node 18+. Claude Agent SDK as a peer (you already have it). Ships ESM + CJS with TypeScript types.

License

MIT