sentry-setup-ai-monitoring

by sentry

Setup Sentry AI Agent Monitoring in any project. Use when asked to monitor LLM calls, track AI agents, or instrument OpenAI/Anthropic/Vercel… npx skills add https://github.com/getsentry/sentry-for-ai --skill sentry-setup-ai-monitoring Download ZIPGitHub232

All Skills > Feature Setup > AI Monitoring

Invoke This Skill When

• User asks to "monitor AI/LLM calls" or "track OpenAI/Anthropic usage"

• User wants "AI observability" or "agent monitoring"

• User asks about token usage, model latency, or AI costs

Important: The SDK versions, API names, and code samples below are examples. Always verify against docs.sentry.io before implementing, as APIs and minimum versions may have changed.

Data Capture Warning

Prompt and output recording captures user content that is likely PII. In JavaScript, genAI input/output capture is on by default (governed by dataCollection.genAI); in Python it is enabled via send_default_pii=True. Before relying on this capture (or per-integration overrides — recordInputs/recordOutputs in JS, include_prompts in Python), confirm:

• The application's privacy policy permits capturing user prompts and model responses

• Captured data complies with applicable regulations (GDPR, CCPA, etc.)

• Sentry data retention settings are appropriate for the sensitivity of the data

Ask the user whether they want prompt/output capture enabled. Do not enable prompt/output capture without explicit confirmation. Use tracesSampleRate: 1.0 only in development; in production, use a lower value or a tracesSampler function.

Detection First

Always detect installed AI SDKs before configuring:

# JavaScript
grep -E '"(openai|@anthropic-ai/sdk|ai|@langchain|@google/genai)"' package.json

# Python
grep -E '(openai|anthropic|langchain|huggingface)' requirements.txt pyproject.toml 2>/dev/null

Sampling Check

After detecting AI SDKs, check the current sampling configuration:

# JavaScript
grep -E 'tracesSampleRate|tracesSampler' sentry.*.config.* instrument.* src/instrument.* app/instrument.* 2>/dev/null

# Python
grep -E 'traces_sample_rate|traces_sampler' *.py **/*.py 2>/dev/null

If tracesSampleRate / traces_sample_rate is below 1.0 AND no tracesSampler / traces_sampler is configured:

Ask the user:

"Your current sample rate is {rate}. Agent runs are sampled as complete span trees — if the root span is dropped, all child gen_ai spans are lost. For full AI visibility, gen_ai-related transactions should be sampled at 100%. Would you like me to set up a tracesSampler that keeps AI traces at 100% while sampling other traffic at your current rate?"

If user confirms, read ${SKILL_ROOT}/references/sampling.md for implementation patterns.

Supported SDKs

JavaScript

Package Integration Min Sentry SDK Auto? openai openAIIntegration() 10.53.0 Yes @anthropic-ai/sdk anthropicAIIntegration() 10.53.0 Yes ai (Vercel) vercelAIIntegration() 10.53.0 Yes* @langchain/* langChainIntegration() 10.53.0 Yes @langchain/langgraph langGraphIntegration() 10.53.0 Yes @google/genai googleGenAIIntegration() 10.53.0 Yes

*Vercel AI: 10.53.0+ required. Requires experimental_telemetry per-call.

Python

Integrations auto-enable when the AI package is installed — no explicit registration needed:

Package Auto? Notes openai Yes Includes OpenAI Agents SDK anthropic Yes langchain / langgraph Yes huggingface_hub Yes google-genai Yes pydantic-ai Yes litellm No Requires explicit integration mcp (Model Context Protocol) Yes

Manual Instrumentation

Use when no supported SDK is detected. Follow the canonical Sentry Conventions for gen_ai.* attributes — the JS docs may lag behind; do not set attributes marked deprecated in the conventions.

Span Types

op Span name pattern Purpose gen_ai.{operation} (e.g. gen_ai.chat, gen_ai.request) {operation} {model} (e.g. chat gpt-4o) Individual LLM call gen_ai.invoke_agent invoke_agent {agent_name} Agent execution lifecycle gen_ai.execute_tool execute_tool {tool_name} Tool/function call gen_ai.handoff handoff from {source} to {target} Agent-to-agent transition

For LLM-call spans, the op follows the pattern gen_ai.{gen_ai.operation.name} — use gen_ai.chat, gen_ai.embeddings, gen_ai.generate_content, or gen_ai.text_completion where the operation is known. Span attributes only accept primitives; arrays/objects must be JSON-stringified.

Example (JavaScript)

const inputMessages = [
 { role: "user", parts: [{ type: "text", content: "Tell me a joke" }] },
];

await Sentry.startSpan({
 op: "gen_ai.chat",
 name: "chat gpt-4o",
 attributes: {
 "gen_ai.request.model": "gpt-4o",
 "gen_ai.operation.name": "chat",
 "gen_ai.input.messages": JSON.stringify(inputMessages),
 },
}, async (span) => {
 const result = await llmClient.complete(inputMessages);

 const outputMessages = [
 {
 role: "assistant",
 parts: [
 // Thinking/reasoning content goes in a `reasoning` part, NOT a `text` part.
 // Sentry surfaces it separately and filters it out of the Conversations view.
 { type: "reasoning", content: result.reasoning },
 { type: "text", content: result.text },
 ],
 finish_reason: result.finishReason,
 },
 ];
 span.setAttribute("gen_ai.output.messages", JSON.stringify(outputMessages));
 span.setAttribute("gen_ai.usage.input_tokens", result.inputTokens);
 span.setAttribute("gen_ai.usage.output_tokens", result.outputTokens);
 return result;
});

Key Attributes

Common (all AI spans):

Attribute Required Description gen_ai.request.model Yes Model identifier (e.g., gpt-4o, claude-sonnet-4-6) gen_ai.operation.name No Operation label (chat, embeddings, invoke_agent, execute_tool, handoff, etc.) gen_ai.agent.name No Agent name (set on agent and tool spans)

Model config (LLM call spans):

Attribute Description gen_ai.request.reasoning_effort Reasoning effort level for reasoning models (e.g., low, medium, high). Supported values vary by provider.

Request / response content (PII — enable only after confirming; see Data Capture Warning above):

Attribute Description gen_ai.input.messages JSON-stringified array of input messages. Each item uses {role, parts} where parts is [{type, content}]; role is "user", "assistant", "tool", or "system". Common part types: "text", "reasoning", "tool_call", "tool_call_response" gen_ai.output.messages JSON-stringified array of response messages (text + tool calls), same shape as inputs

Thinking / reasoning messages: Models with extended thinking (Anthropic thinking blocks, Gemini thought, DeepSeek reasoning_content) produce internal reasoning that isn't part of the user-visible reply. Represent it as a reasoning part inside the assistant message — {"type": "reasoning", "content": "..."} — alongside the user-facing text part. Sentry surfaces reasoning parts separately and filters them out of the user-facing Conversations view, so do not fold thinking into a text part. When previous thinking is fed back into a multi-turn request, include the same reasoning parts in the assistant messages within gen_ai.input.messages. Record reasoning token counts via gen_ai.usage.output_tokens.reasoning (a subset of gen_ai.usage.output_tokens). | gen_ai.system_instructions | System prompt passed to the model | | gen_ai.tool.definitions | JSON-stringified list of tools available to the model |

Token usage:

Attribute Description gen_ai.usage.input_tokens Total input tokens — includes cached tokens gen_ai.usage.input_tokens.cached Subset of input tokens served from cache gen_ai.usage.input_tokens.cache_write Tokens written to cache while processing input gen_ai.usage.output_tokens Total output tokens — includes reasoning tokens gen_ai.usage.output_tokens.reasoning Subset of output tokens used for reasoning gen_ai.usage.total_tokens Sum of input + output tokens

Tool spans (gen_ai.execute_tool):

Attribute Description gen_ai.tool.name Tool identifier gen_ai.tool.description Human-readable tool description gen_ai.tool.call.arguments JSON-stringified tool arguments gen_ai.tool.call.result JSON-stringified tool result

Token Usage and Cost Calculation

Sentry uses token attributes to calculate model costs. Cached and reasoning tokens are subsets, not separate counts — gen_ai.usage.input_tokens already includes gen_ai.usage.input_tokens.cached, and gen_ai.usage.output_tokens already includes gen_ai.usage.output_tokens.reasoning.

Sentry subtracts the cached/reasoning counts from the totals to compute the uncached/non-reasoning portion. Reporting a cached or reasoning count greater than its total produces negative costs in the dashboard.

Example — 100 input tokens total, 90 served from cache:

• Correct: input_tokens = 100, input_tokens.cached = 90

• Wrong: input_tokens = 10, input_tokens.cached = 90 (cached larger than total → negative cost)

The same rule applies to gen_ai.usage.output_tokens vs. gen_ai.usage.output_tokens.reasoning.

Verification

After configuring, make an LLM call and check the Sentry Traces dashboard. AI spans appear with gen_ai.* operations showing model, token counts, and latency.

Conversations

Conversations gives a readable, chat-style view of past sessions with your AI agent. It groups spans by gen_ai.conversation.id — so whether a user talked across multiple traces or multiple conversations happened inside one trace, you get a timeline of every message, tool call, and response.

When the user asks for AI monitoring setup, proactively mention this requirement if the app has multi-turn chats. Without a conversation ID, the agent-monitoring spans still work, but the Conversations view cannot group the session correctly.

Find it at Explore > Conversations in Sentry.

Prerequisites for Conversations

• Tracing enabled with tracesSampleRate > 0

• Gen AI span streaming is on by default — streamGenAiSpans defaults to true since JS SDK 10.61.0 and stream_gen_ai_spans defaults to True since Python SDK 2.64.0. This sends AI spans as standalone items, so spans with large inputs/outputs don't hit transaction payload size limits and get dropped. (The options are available since JS 10.53.0 / Python 2.60.0 if you need to set them explicitly on older SDKs.)

• Input and output capture enabled — Conversations reconstructs the chat from gen_ai.input.messages and gen_ai.output.messages attributes. In JS this is on by default (via dataCollection); in Python, set send_default_pii=True. Without it, conversations appear empty.

Setting a Conversation ID

Some integrations (OpenAI Agents SDK for Python, OpenAI SDK for Node) infer the conversation ID automatically. For all others, set it manually.

JavaScript

import * as Sentry from "@sentry/node"; // or @sentry/nextjs, @sentry/nestjs, etc.

// Set at the start of a conversation
Sentry.setConversationId("conv_abc123");

// All subsequent AI calls carry gen_ai.conversation.id: "conv_abc123"
await openai.chat.completions.create({
 model: "gpt-5.5",
 messages: [{ role: "user", content: "Hello" }],
});

Python

import sentry_sdk.ai

# Set at the start of a conversation
sentry_sdk.ai.set_conversation_id("conv_abc123")

# All subsequent AI calls carry gen_ai.conversation.id = "conv_abc123"

Some integrations infer the conversation ID automatically. For example, the Python OpenAI integration picks it up when you use the conversation parameter:

import openai
import sentry_sdk

sentry_sdk.init(...)

conversation = openai.conversations.create()
response = openai.responses.create(
 model="gpt-5.4",
 input=[{"role": "user", "content": "What are the 5 Ds of dodgeball?"}],
 conversation=conversation.id # automatically sets gen_ai.conversation.id
)

User Attribution

The Conversations view shows a User column. To populate it, call setUser / set_user once per request or session, before any AI calls:

JavaScript

import * as Sentry from "@sentry/node"; // or @sentry/nextjs, @sentry/nestjs, etc.

Sentry.setUser({ id: "user_123", email: "[email protected]", username: "jane" });

Python

import sentry_sdk

sentry_sdk.set_user({"id": "user_123", "email": "[email protected]", "username": "jane"})

Any of id, email, or username is sufficient — Conversations will display whichever fields are present.

Conversations vs Traces

These are independent concepts:

• A single conversation can span multiple traces (e.g., user refreshes the page mid-conversation — new trace, same conversation ID)

• A single trace can contain spans from different conversations (e.g., user starts a new chat without refreshing)