IndraTrace SDK

OpenTelemetry-native observability SDK for the IndraTrace platform — one-line instrumentation for web apps and AI agents: traces, logs, metrics, and model-call token usage.

pip install indratrace

from indratrace import init_observability

init_observability(api_key="it_live_...")

That is the whole integration. The API key is the only thing you configure — it identifies your product, its environment and your workspace, all decided when you register the product and copy its key. There is nothing else to set, and no endpoint to point at.

Two words to know up front:

• A span is one timed step — a web request, a database call, one model call.
• A trace is the full story of one request — its spans stacked on a timeline, so you can see where the time went and what called what.

init_observability() ships traces, logs, and metrics; trace_agent / trace_tool wrap your agents and tools; and model spans carry exact, provider-reported token counts when the GenAI extras are installed. Token counts are recorded raw — the SDK never computes cost; the platform derives it at query time.

import logging

from fastapi import FastAPI

from indratrace import init_observability, trace_agent, trace_tool

# Once, at app startup. The key is the whole configuration.
init_observability(api_key="it_live_...")

app = FastAPI()  # every HTTP request becomes a span, automatically


@trace_tool  # a span per tool call
async def risk_score(vendor: str) -> int:
    logging.getLogger(__name__).info("scoring %s", vendor)  # ships with trace context
    return len(vendor)


@trace_agent("compliance-checker")  # a span wrapping the whole agent request
async def run(query: str) -> int:
    return await risk_score(query)

The import logging above is just Python's built-in logging — nothing IndraTrace-specific, and not required. Whatever your app already logs ships automatically (see Configuration); the line is there to show a log call picking up its span's trace context.

Both decorators work on sync and async functions. They are transparent: a tool that raises gets its span marked ERROR with the exception recorded, and the exception then propagates to your code unchanged.

What you get, by framework

If your app uses Python's standard logging, your logs are already in IndraTrace — don't use print(). That is true with no extra installed and no configuration: the one init_observability() call bridges logging into the pipeline, and every record emitted inside a span carries that span's trace context. As of 0.6.0 the same is true of Loguru.

Everything else — HTTP server spans, model spans, agent spans — comes from the extras in the table below.

Install exactly what you use — extras are additive, and core stays dependency-clean (OpenTelemetry only):

pip install "indratrace[fastapi,anthropic]"        # a FastAPI app calling Claude
pip install "indratrace[django]"                   # a Django app
pip install "indratrace[flask,openai]"             # a Flask app calling OpenAI

¹ Django: init_observability() must run before Django builds its application object — it works by adding middleware. See Django. ² Flask: if your module does from flask import Flask, add one line — instrument_flask_app(app). See Flask.

Missing an extra is never an error — the SDK skips it silently. If a signal you expected is missing, turn on debug: the startup banner prints enabled or skipped (extra not installed) for every integration above.

Tracing a regular REST API

You don't need to be building an AI agent. For a plain FastAPI service, the one init line is the whole setup — every endpoint is reported as a span, with its status and duration, no per-route code:

from fastapi import FastAPI

from indratrace import init_observability

init_observability(api_key="it_live_...")

app = FastAPI()


@app.get("/orders/{order_id}")            # this endpoint is now a span automatically
def get_order(order_id: str) -> dict:
    return {"id": order_id}

When one endpoint is slow and you want to see inside it — which query, which parser ate the time — wrap that piece in @trace_step. It adds a child span under the request so the timeline shows the breakdown:

from indratrace import trace_step


@trace_step                               # a span named "step load_order"
def load_order(order_id: str) -> dict:
    ...                                   # e.g. a database query
    return {"id": order_id}

@trace_step is the neutral sibling of @trace_tool: same behavior, but for timing ordinary functions (database queries, parsers, validation) where calling them a "tool" would be misleading. Bare or called (@trace_step()), sync or async, exceptions recorded and re-raised unchanged.

Install the extras you need — FastAPI for HTTP auto-instrumentation, and anthropic / openai / gemini / bedrock for model spans with token usage:

pip install "indratrace[fastapi,anthropic,openai,gemini,bedrock]"

Django

pip install "indratrace[django]"

Every request becomes a server span, with no per-view code. Where you call init_observability() matters. The instrumentation works by adding middleware, and Django reads its middleware list once, when it builds the application object — so init has to happen before that. In practice: put it at the top of wsgi.py (and asgi.py, and manage.py if you use runserver), above the get_wsgi_application() call.

# myproject/wsgi.py
import os

from django.core.wsgi import get_wsgi_application

from indratrace import init_observability

os.environ.setdefault("DJANGO_SETTINGS_MODULE", "myproject.settings")

init_observability(api_key="it_live_...")            # BEFORE ↓

application = get_wsgi_application()

Call it after get_wsgi_application() and you get logs and model spans, but no HTTP spans — and no error saying so. The reason: get_wsgi_application() reads settings.MIDDLEWARE and freezes a middleware chain from it, so a middleware added later is simply never in the chain your server actually runs. If your HTTP spans are missing, check this first.

A wrinkle worth knowing if you write tests: Django's test Client builds a fresh handler per request, so it re-reads settings.MIDDLEWARE every time and will happily produce spans even when init ran too late. Only a real WSGI/ASGI server exposes the mistake — so trust your staging environment here, not a passing test.

Flask

pip install "indratrace[flask]"

Every request becomes a server span. One caveat, and it bites the most common import style. The instrumentation works by replacing the flask.Flask class, so an app built from a Flask name that was imported before init_observability() ran is left uninstrumented — silently. Since from flask import Flask sits at the top of the file and init runs below it, that is the usual case. Add one line:

from flask import Flask

from indratrace import init_observability, instrument_flask_app

init_observability(api_key="it_live_...")

app = Flask(__name__)
instrument_flask_app(app)          # ← now every request is a span


@app.get("/orders/<order_id>")
def get_order(order_id: str):
    return {"id": order_id}

instrument_flask_app(app) is safe to call twice, never raises, and is a no-op if the flask extra isn't installed. (If you construct the app as flask.Flask(__name__) — looking the name up on the module instead of importing the class — the extra line isn't needed. The explicit call works either way, so when in doubt, keep it.)

Loguru

Nothing to install and nothing to configure — as of 0.6.0, if loguru is importable, init_observability() bridges it automatically:

from loguru import logger

from indratrace import init_observability

init_observability(api_key="it_live_...")

logger.info("this ships to IndraTrace")          # INFO and above
logger.exception("so does this, with its stack trace")

Records at INFO and above are exported (DEBUG stays local — it would be a firehose), severities are preserved, and a line logged inside a span carries that span's trace context, exactly like a stdlib one. Your own loguru sinks are untouched: console output looks the same as it always did, and an app that uses loguru and stdlib logging gets each record exported exactly once.

If you configure loguru after init

logger.remove() — the idiomatic way to drop loguru's default stderr sink — takes every sink with it, including ours. So an app that reconfigures loguru after init_observability() silently unbridges itself. Put the bridge back with bridge_loguru():

from loguru import logger

from indratrace import bridge_loguru, init_observability

init_observability(api_key="it_live_...")

logger.remove()                    # your own setup — drops our sink too
logger.add("app.log", level="INFO")

bridge_loguru()                    # ← put the bridge back; logs ship again

Call it any time after init; it's idempotent, so calling it twice does not double-export. It returns False (and does nothing) if loguru isn't installed or init_observability() never ran.

The simplest way to avoid the whole issue is to configure loguru before init_observability(), in which case there is nothing to re-add.

Claude Agent SDK — zero instrumentation

If your app is built on Anthropic's claude-agent-sdk, the single init_observability() call traces the whole agent loop — with no decorators anywhere. Install the extra, init once, and every query() / ClaudeSDKClient run produces:

• an agent span for the run,
• a turn span for each step of the loop, each with that turn's model and exact token usage (input, output, and cache tokens),
• a tool span for every tool the agent calls — including MCP tools, tagged with the MCP server name — with an error status if the tool failed.

They arrive as one nested trace, so you can see exactly what the agent did, how many turns it took, which tools it called, and how many tokens each step spent.

pip install "indratrace[claude-agent-sdk]"

from claude_agent_sdk import query, ClaudeAgentOptions

from indratrace import init_observability

init_observability(api_key="it_live_...")   # once, at startup

# No decorators. This whole run is traced — agent → turns → tools → tokens.
async for message in query(prompt="Summarize today's incidents and file a ticket"):
    print(message)

That's it — there is nothing else to add. It works the same with the stateful ClaudeSDKClient (each receive_response() is one traced run), it nests inside a @trace_agent if you have one, it picks up session(...) context, and an early-abandoned stream never leaves a span dangling. Token counts are stored raw; the SDK never computes cost.

Under the hood the Agent SDK runs the agent loop in a subprocess CLI, so the usage is read straight off the SDK's own messages — but it lands under the same gen_ai.usage.* names as any other model span.

Token usage from model calls

With the anthropic, openai, gemini, or bedrock extra installed, every provider call made after init_observability() produces a model span carrying the exact, provider-reported token counts (gen_ai.usage.input_tokens / gen_ai.usage.output_tokens), nested under whatever agent/tool span is active. No wrapper, no config — just call the provider as you already do:

import anthropic

from indratrace import init_observability, trace_agent, trace_tool

init_observability(api_key="it_live_...")
client = anthropic.Anthropic()


@trace_tool
def summarize(doc: str) -> str:
    msg = client.messages.create(               # model span with token counts,
        model="claude-haiku-4-5",               # a child of this tool span
        max_tokens=256,
        messages=[{"role": "user", "content": doc}],
    )
    return msg.content[0].text


@trace_agent("summarizer")
def run(doc: str) -> str:
    return summarize(doc)

Streaming calls are captured too — usage lands on the span from the final streamed event.

For a provider the SDK does not auto-instrument, stamp the counts yourself from inside a span with record_llm_usage:

from indratrace import record_llm_usage

record_llm_usage(
    model="some-model-v2",
    input_tokens=resp.usage.input,
    output_tokens=resp.usage.output,
    system="acme-ai",
)

Token counts are stored raw — the SDK never computes cost; the platform derives it at query time from a price table.

Capturing prompt & completion text

By default, model spans carry token counts but not the prompt or completion text — because prompts often contain customer data. Turn the text on when you want to see exactly what was sent and returned (the usual case in dev and staging, off in production):

init_observability(api_key="it_live_...", capture_content=True)

Or set INDRATRACE_CAPTURE_CONTENT=true in the environment (an explicit capture_content= argument wins over it). When on, the prompt lands on the model span under gen_ai.input.messages and the completion under gen_ai.output.messages. This flag only gates the text — token counts are captured either way.

Session & user context

Wrap a conversation in session(...) and every span started inside it — your agent/tool spans, the FastAPI HTTP span, and the GenAI model spans — carries session.id and/or user.id. No per-call wiring: the ids ride OTel baggage and a span processor stamps them at span start.

from indratrace import session

with session(session_id="conversation-42", user_id="u-1001"):
    answer = run(query)          # every span here is tagged with both ids

It works across async/await and threads, and nests — an inner session(user_id=...) overrides only user.id and keeps the outer session.id. For middleware that can't bracket a with (it tags on request-in and untags on request-out, in separate callbacks), call it imperatively and keep the handle:

handle = session(session_id=request.headers["x-session-id"])
try:
    ...                          # dispatch the request
finally:
    handle.detach()             # or handle.close(); restores the prior context

Feedback (👍 / 👎)

Tie a user's thumbs-up/down back to the trace that produced the answer. Capture the trace id at answer time with current_trace_id(), hand it back to the caller, and record the score whenever the user reacts — often minutes later, out of band:

from indratrace import current_trace_id, record_feedback

@trace_agent("assistant")
def answer(query: str) -> dict:
    text = run(query)
    return {"answer": text, "trace_id": current_trace_id()}  # store this id

# later, when the user clicks 👍
record_feedback(1, comment="spot on", trace_id=stored_trace_id)

score is any number — the convention is 1 for positive, 0/-1 for negative, but any scale (e.g. 1–5) works. If you omit trace_id, the current trace's id is used when you're inside one. record_feedback emits a short feedback span carrying feedback.score, the optional feedback.comment, and feedback.trace_id, which the platform joins back to the original trace. Called inside session(...), the feedback span carries the session/user ids too.

Configuration

There is one required setting, and it is the API key:

init_observability(api_key="it_live_...")

export INDRATRACE_API_KEY="it_live_..."   # or set it in the environment

An explicit api_key= wins over the environment. If neither is set, init_observability() raises IndraTraceConfigError and says so — it will not quietly run un-instrumented, because a missing key is the one mistake you cannot see from the outside.

Where the key comes from: register the product in your IndraTrace workspace, choose its environment there, and copy the key it shows you. The key is shown once, at creation. A key belongs to one product in one environment — a staging deploy and a production deploy of the same product hold different keys, and that is the only difference between them.

Everything else is optional and keyword-only:

service_name is for one product with several deployables — an API and a worker, say. It labels the deployable; it does not name your product, which the key already does.

init_observability() also takes instrument_http=False to turn off web-framework auto-instrumentation entirely (it's on by default, and an absent extra is already a no-op). Before 0.6.0 this argument was called instrument_fastapi; the old name still works and now gates all three frameworks.

Upgrading from 0.6? product, env, endpoint and ingest_key were removed in 1.0 — the key decides all of them. Passing one raises an error that names it and tells you what to write instead, and INDRATRACE_PRODUCT / INDRATRACE_ENV / INDRATRACE_KEY are ignored with a warning. See the CHANGELOG for a before/after snippet.

Your existing logging calls ship automatically once your app is at INFO — the usual case under basicConfig(level=INFO), uvicorn, or gunicorn. The SDK does not change your root logger's level on its own; if your app never configured logging (so it sits at the stdlib default of WARNING), pass log_level="INFO" to opt in:

init_observability(api_key="it_live_...", log_level="INFO")

Loguru needs none of this: its own level gates its records, and the bridge takes everything at INFO and above regardless of the stdlib root level (see Loguru).

The SDK never raises into your app after it is wired: if IndraTrace is unreachable or rejects the key, it logs and drops, never blocks a request and never raises. The decorators hold to that too — they run your function even when init_observability() was never called.

The one exception is init_observability() itself, which raises IndraTraceConfigError when there is no API key (or when you pass a parameter that 1.0 removed). That is a startup mistake, on a line you are looking at — and letting it through would only turn into telemetry silently rejected in production, which is the failure this replaced.

Still waiting for your first span?

You called init_observability(), your app is serving traffic, and the dashboard is empty. There are only three things it can be — work down the list.

1. Is INDRATRACE_API_KEY actually set in the process that is running?

python -c "import os; print(os.getenv('INDRATRACE_API_KEY') or 'NOT SET')"

Run that the same way you run your app — same shell, same container, same systemd unit. A .env file your process manager never loads, or a key exported in your terminal but not in the container, is the single most common cause. If the key is missing entirely, init_observability() raises rather than starting un-instrumented — so if your app started at all, a key was found somewhere. The question is whether it is the right one.

2. Is it the key for the product you expect?

A key belongs to one product in one environment. Your telemetry is arriving under whichever product that key names — so a staging key in a production deploy means the data is landing in staging, not nowhere. Open Products in your IndraTrace workspace and check the environment you are looking at matches the key the process is holding. (Keys are shown once at creation; if you cannot tell which one a deploy has, rotate it and set the new one deliberately.)

3. Turn on debug and read the export lines.

export INDRATRACE_DEBUG=1

or pass debug=True. It prints a startup banner and turns silent drops into visible log lines (details below). Two lines answer almost everything:

• export ok — the telemetry left your process and was accepted. If the dashboard is still empty, you are looking at the wrong product or environment; go back to step 2.
• export FAILED — it did not, and the line names the HTTP status the gateway answered with. A 401 means the key was rejected (revoked, or a typo in the value); no HTTP response at all means the gateway was not reachable from where your app runs; any other status means it was reached and refused the batch.

If all three check out and you still see nothing, the remaining possibility is placement, not configuration: init_observability() ran too late for Django or after your Flask app object was built — see Flask. The banner cannot detect either.

Turning on debug

Because the SDK is fail-silent — it never raises or blocks your app once it is wired — a missing extra or an unreachable gateway can leave your dashboard empty with no obvious clue why. Pass debug=True to make those failures audible:

init_observability(api_key="it_live_...", debug=True)

or set INDRATRACE_DEBUG=1 in the environment. It prints a startup banner and turns silent drops into visible log lines — without changing behavior; your app still never sees an exception from the SDK.

indratrace [INFO] indratrace initialized: service=my-app endpoint=http://localhost:8088
indratrace [DEBUG] IndraTrace SDK v1.0.0 initialized
indratrace [DEBUG]   service=my-app version=1.4.2
indratrace [DEBUG]   endpoint=http://localhost:8088 (traces=http://localhost:8088/v1/traces)
indratrace [DEBUG]   api_key=set capture_content=off
indratrace [DEBUG]   identity: product, deployment.environment and tenant.id are stamped by IndraTrace from your API key
indratrace [DEBUG]   signals: traces + logs + metrics (OTLP/HTTP, batched)
indratrace [DEBUG]   http[fastapi]: skipped (extra not installed)
indratrace [DEBUG]   http[django]: enabled
indratrace [DEBUG]   http[flask]: skipped (extra not installed)
indratrace [DEBUG]   loguru: enabled
indratrace [DEBUG]   genai[anthropic]: enabled
indratrace [DEBUG]   claude-agent-sdk: skipped (extra not installed)
indratrace [WARNING] indratrace: traces export FAILED (FAILURE) — HTTP 401: the gateway rejected your API key. Check the value of INDRATRACE_API_KEY in the process that is running; keys are shown once when the product is created.

Read it top to bottom:

• The identity line is there because the banner reports no product and no environment — deliberately. Your key decides both, and IndraTrace stamps them when the telemetry lands, so there is nothing here for the SDK to report.
• Each integration line says enabled or skipped (reason). skipped (extra not installed) means you need the extra, e.g. pip install "indratrace[anthropic]". The http[…], loguru, genai[…], and claude-agent-sdk lines cover every row of the support matrix.
• An export FAILED line means the SDK built fine but IndraTrace didn't accept the data. It carries the HTTP status the gateway answered with: 401 is a rejected key (the sample above), any other status means the gateway was reached and refused the batch, and no HTTP response at all means it was not reachable from where your app runs.

One thing the banner cannot tell you: whether init_observability() ran early enough. http[django]: enabled means the middleware was installed, but if you called init after get_wsgi_application() it went into a chain Django had already built, and you'll still see no HTTP spans — see Django. Likewise http[flask]: enabled doesn't guarantee your app object was caught; see Flask.

The debug lines go to your console only (they're never shipped to the platform), and debug defaults to off, so production stays quiet. Turn it off once you've found the problem.

Contributing & community

• CONTRIBUTING.md — dev setup, the test harness, and PR expectations.
• SECURITY.md — how to report a vulnerability privately.
• CODE_OF_CONDUCT.md — the Contributor Covenant we follow.
• Found a bug or want a feature? Open an issue.

Built by Indrasol.