:hedgehog: PostHog is the leading platform for building self-driving products. Our developer tools – AI observability, analytics, session replay, flags, experiments, error tracking, logs, and more – capture all the context agents need to diagnose problems, uncover opportunities, and ship fixes. Steer it all from Slack, web, desktop, or the MCP.
Python ab-testingai-analyticsanalyticscdpdata-warehouseexperimentsfeature-flagsjavascriptproduct-analyticspythonreactsession-replaysurveystypescriptweb-analytics
feat(mcp-analytics): route llm calls through the ai gateway (#128189)
Jonathan Mieloo579d3bf48b5a
diff --git a/products/mcp_analytics/backend/failure_classification.py b/products/mcp_analytics/backend/failure_classification.pyindex 15d50aa6a508..afc3cdce6d37 100644--- a/products/mcp_analytics/backend/failure_classification.py+++ b/products/mcp_analytics/backend/failure_classification.py@@ -8,15 +8,13 @@ pattern in ``intent_generation.py``.import refrom enum import StrEnum-from django.conf import settings-import openaiimport structlog-import posthoganalytics-from posthoganalytics.ai.openai import OpenAI+from openai import OpenAIfrom pydantic import BaseModel, ValidationErrorfrom posthog.clickhouse.query_tagging import Feature, Product, tags_context+from posthog.llm.gateway_client import build_openai_client, resolve_ai_gateway_configfrom posthog.models.team.team import Teamlogger = structlog.get_logger(__name__)@@ -160,7 +158,6 @@ def _build_batch_prompt(messages: list[str]) -> str:def _call_llm(client: OpenAI, messages: list[str], team: Team) -> ClassificationBatch:response = client.chat.completions.create(model=FAILURE_CLASS_MODEL,- temperature=0,timeout=30,messages=[{"role": "system", "content": SYSTEM_PROMPT},@@ -168,7 +165,6 @@ def _call_llm(client: OpenAI, messages: list[str], team: Team) -> Classification],response_format={"type": "json_object"},user=f"team/{team.id}/mcp-analytics-failure-classifier",- posthog_properties={"ai_product": "mcp_analytics", "ai_feature": "failure-classification"},)content = response.choices[0].message.content if response.choices else Noneif not content:@@ -216,7 +212,14 @@ def classify_fingerprints(messages_by_fingerprint: dict[str, str], team: Team) -if not team.organization.is_ai_data_processing_approved:raise FailureClassificationUnavailable("AI data processing is not approved for this organization")- client = OpenAI(posthog_client=posthoganalytics.default_client, base_url=settings.OPENAI_BASE_URL)+ if resolve_ai_gateway_config() is None:+ raise FailureClassificationUnavailable("AI gateway is not configured")+ client = build_openai_client(+ "django",+ ai_product="mcp_analytics",+ properties={"ai_feature": "failure-classification"},+ distinct_id=f"team/{team.id}/mcp-analytics-failure-classifier",+ )items = list(messages_by_fingerprint.items())result: dict[str, str] = {}with tags_context(diff --git a/products/mcp_analytics/backend/intent_generation.py b/products/mcp_analytics/backend/intent_generation.pyindex f3caca1f0188..8bc044ff4ef0 100644--- a/products/mcp_analytics/backend/intent_generation.py+++ b/products/mcp_analytics/backend/intent_generation.py@@ -7,15 +7,11 @@ generation only — caching and persistence live in ``logic.generate_session_intfrom datetime import datetime, timedelta-from django.conf import settingsfrom django.utils import timezoneimport openaiimport structlog-import posthoganalyticsfrom google.genai.types import GenerateContentConfig-from posthoganalytics.ai.gemini import genai-from posthoganalytics.ai.openai import OpenAIfrom pydantic import BaseModel, Field, ValidationErrorfrom posthog.hogql import ast@@ -23,6 +19,7 @@ from posthog.hogql.parser import parse_selectfrom posthog.hogql.query import execute_hogql_queryfrom posthog.clickhouse.query_tagging import Feature, Product, tags_context+from posthog.llm.gateway_client import build_gemini_client, build_openai_client, resolve_ai_gateway_configfrom posthog.models.team.team import Teamfrom products.mcp_analytics.backend.constants import MCP_TOOL_CALL_EVENT@@ -92,11 +89,11 @@ def _build_user_prompt(intents: list[str]) -> str:return f"Per-tool-call intents (chronological):\n{numbered}\n\nSummarise the agent's overall goal in at most two short, concrete sentences."-def _ensure_llm_available(team: Team, *, api_key: str, key_name: str) -> None:+def _ensure_llm_available(team: Team, *, configured: bool) -> None:"""Intents are user-authored text, so sending them to the LLM requires theorganization's AI data processing consent, same as every other AI-backed flow."""- if not api_key:- raise IntentGenerationUnavailable(f"{key_name} is not configured")+ if not configured:+ raise IntentGenerationUnavailable("AI gateway is not configured")if not team.organization.is_ai_data_processing_approved:raise IntentGenerationUnavailable("AI data processing is not approved for this organization")@@ -108,9 +105,15 @@ def summarize_intents(intents: list[str], team: Team) -> str:has not approved AI data processing, or the request fails, so the endpoint can answerwith a clean 503 rather than a 500."""- _ensure_llm_available(team, api_key=settings.OPENAI_API_KEY, key_name="OPENAI_API_KEY")-- client = OpenAI(posthog_client=posthoganalytics.default_client, base_url=settings.OPENAI_BASE_URL)+ _ensure_llm_available(team, configured=resolve_ai_gateway_config() is not None)++ distinct_id = f"team/{team.id}/mcp-analytics-session-intent"+ client = build_openai_client(+ "django",+ ai_product="mcp_analytics",+ properties={"ai_feature": "session-intent-generation"},+ distinct_id=distinct_id,+ )try:response = client.chat.completions.create(model=SESSION_INTENT_MODEL,@@ -121,8 +124,7 @@ def summarize_intents(intents: list[str], team: Team) -> str:{"role": "system", "content": SYSTEM_PROMPT},{"role": "user", "content": _build_user_prompt(intents)},],- user=f"team/{team.id}/mcp-analytics-session-intent",- posthog_properties={"ai_product": "mcp_analytics", "ai_feature": "session-intent-generation"},+ user=distinct_id,)except openai.OpenAIError as e:raise IntentGenerationUnavailable("LLM request failed") from e@@ -235,13 +237,14 @@ def summarize_project_intents(intents: list[tuple[str, str]], team: Team) -> Inthas not approved AI data processing, or the request fails, so the endpoint can answerwith a clean 503 rather than a 500."""- _ensure_llm_available(team, api_key=settings.GEMINI_API_KEY, key_name="GEMINI_API_KEY")-- client = genai.Client(- api_key=settings.GEMINI_API_KEY,- posthog_client=posthoganalytics.default_client,- http_options={"timeout": DIGEST_TIMEOUT_MS},+ client = build_gemini_client(+ ai_product="mcp_analytics",+ properties={"ai_feature": "activity-intent-digest"},+ distinct_id=f"team/{team.id}/mcp-analytics-intent-digest",+ timeout_ms=DIGEST_TIMEOUT_MS,)+ _ensure_llm_available(team, configured=client is not None)+ assert client is not Noneconfig = GenerateContentConfig(system_instruction=DIGEST_SYSTEM_PROMPT,response_mime_type="application/json",@@ -253,9 +256,6 @@ def summarize_project_intents(intents: list[tuple[str, str]], team: Team) -> Intmodel=DIGEST_THEMES_MODEL,contents=_build_digest_prompt(intents),config=config,- posthog_distinct_id=f"team/{team.id}/mcp-analytics-intent-digest",- posthog_properties={"ai_product": "mcp_analytics", "ai_feature": "activity-intent-digest"},- posthog_groups={"project": str(team.id)},)except Exception as e:# Broad: the Gemini client surfaces transport, timeout, and API errors as unrelated types,diff --git a/products/mcp_analytics/backend/tests/test_api.py b/products/mcp_analytics/backend/tests/test_api.pyindex d295d1aeb823..6a6af850ca11 100644--- a/products/mcp_analytics/backend/tests/test_api.py+++ b/products/mcp_analytics/backend/tests/test_api.py@@ -766,8 +766,8 @@ class TestResolveIntentThemes(SimpleTestCase):class TestLLMConsentGate(APIBaseTest):@parameterized.expand([- ("session_summary", intent_generation.summarize_intents, "OpenAI"),- ("project_digest", intent_generation.summarize_project_intents, "genai"),+ ("session_summary", intent_generation.summarize_intents, "build_openai_client"),+ ("project_digest", intent_generation.summarize_project_intents, "build_gemini_client"),])def test_refuses_without_ai_data_processing_consent(self, _name, summarize, client_attr: str) -> None:@@ -775,12 +775,12 @@ class TestLLMConsentGate(APIBaseTest):self.organization.save()with (- self.settings(OPENAI_API_KEY="sk-test", GEMINI_API_KEY="gem-test"),+ self.settings(AI_GATEWAY_URL="https://gateway.example.com/v1", AI_GATEWAY_API_KEY="phs_test"),patch.object(intent_generation, client_attr) as mock_client,self.assertRaises(contracts.IntentGenerationUnavailable),):summarize([("find the signups funnel", "query_run")], self.team)- mock_client.assert_not_called()+ assert mock_client.return_value.mock_calls == []class TestActivityOverview(_MCPAnalyticsTeamScopedTestMixin, ClickhouseTestMixin, APIBaseTest):diff --git a/products/mcp_analytics/backend/tests/test_failure_classification.py b/products/mcp_analytics/backend/tests/test_failure_classification.pyindex 837a13090ab2..60b3cdd8c249 100644--- a/products/mcp_analytics/backend/tests/test_failure_classification.py+++ b/products/mcp_analytics/backend/tests/test_failure_classification.py@@ -9,6 +9,8 @@ import jsonfrom posthog.test.base import BaseTestfrom unittest.mock import MagicMock, patch+from django.test import override_settings+from parameterized import parameterizedfrom products.mcp_analytics.backend.failure_classification import (@@ -68,6 +70,7 @@ def _classification_payload(fingerprints: list[str], failure_class: str) -> str:)+@override_settings(AI_GATEWAY_URL="https://gateway.example.com/v1", AI_GATEWAY_API_KEY="phs_test")class TestClassifyFingerprints(BaseTest):def setUp(self) -> None:super().setUp()@@ -77,7 +80,7 @@ class TestClassifyFingerprints(BaseTest):def test_raises_without_ai_processing_consent(self) -> None:self.organization.is_ai_data_processing_approved = Falseself.organization.save()- with patch("products.mcp_analytics.backend.failure_classification.OpenAI") as mock_openai_cls:+ with patch("products.mcp_analytics.backend.failure_classification.build_openai_client") as mock_openai_cls:with self.assertRaises(FailureClassificationUnavailable):classify_fingerprints({"a fingerprint": "a raw message"}, self.team)@@ -96,7 +99,7 @@ class TestClassifyFingerprints(BaseTest):]})- with patch("products.mcp_analytics.backend.failure_classification.OpenAI") as mock_openai_cls:+ with patch("products.mcp_analytics.backend.failure_classification.build_openai_client") as mock_openai_cls:mock_create = mock_openai_cls.return_value.chat.completions.createmock_create.return_value = _openai_response(payload)result = classify_fingerprints(items, self.team)@@ -110,7 +113,7 @@ class TestClassifyFingerprints(BaseTest):def test_batches_at_25_fingerprints(self) -> None:items = {f"fingerprint {i}": f"raw message {i}" for i in range(CLASSIFY_BATCH_SIZE + 1)}keys = list(items)- with patch("products.mcp_analytics.backend.failure_classification.OpenAI") as mock_openai_cls:+ with patch("products.mcp_analytics.backend.failure_classification.build_openai_client") as mock_openai_cls:mock_create = mock_openai_cls.return_value.chat.completions.createmock_create.side_effect = [_openai_response(_classification_payload(keys[:CLASSIFY_BATCH_SIZE], "internal_error")),@@ -137,7 +140,7 @@ class TestClassifyFingerprints(BaseTest):) -> None:items = {"some fingerprint": "some raw message"}payload = json.dumps({"classifications": classifications})- with patch("products.mcp_analytics.backend.failure_classification.OpenAI") as mock_openai_cls:+ with patch("products.mcp_analytics.backend.failure_classification.build_openai_client") as mock_openai_cls:mock_create = mock_openai_cls.return_value.chat.completions.createmock_create.side_effect = [_openai_response(payload), _openai_response(payload)]result = classify_fingerprints(items, self.team)