TruLens
trulens.org
- Free plan
- Open source
- API
- Self-hosted
What TruLens is
TruLens is a free, open-source Python SDK for tracing and evaluating AI agents, RAG applications, MCP apps, and summarization systems. It records per-step latency, inputs, outputs, token use, and cost so teams can inspect why an answer performed poorly. Built-in evaluations cover tool selection, groundedness, context relevance, and correctness. Custom metrics can use Python functions or LLM-as-judge prompts, with judges tuned through rubrics and examples. Evaluations run as traces arrive or in batches over datasets, and a leaderboard compares app versions by metrics, latency, and cost. TruLens uses OpenTelemetry traces and can instrument custom Python applications plus LangGraph and LlamaIndex. It logs to local SQLite by default or to a SQLAlchemy-compatible database URL, and supports a Snowflake connector. Conversation grouping, conversation-level metrics, retrieval APIs, and dashboard views are unavailable with that connector. The project began at TruEra and is maintained by Snowflake.
Who it suits
TruLens suits developers and AI teams evaluating agent, RAG, MCP, or summarization applications. It is appropriate for self-hosted workflows that need tracing, custom metrics, and experiment comparison.
Good to know, and what to watch
- Free open-source Python SDK
- Records latency, tokens, inputs, outputs, and cost
- Supports custom Python and LLM-judge metrics
- Runs streaming or batch evaluations
- Works with OpenTelemetry instrumentation
- SnowflakeConnector lacks conversation features and dashboard views
- Deployment is self-hosted
- Support is directed to a community Slack
Plans
Every plan and priceOpen source
Open-source TruLens library
Open source
The verdict
TruLens provides a self-hosted way to trace AI behavior, define evaluations, and compare application versions. Teams using the Snowflake connector should account for its missing conversation and dashboard capabilities.
What the maker publishes
- Acquisition
- Snowflake announced an agreement on May 22, 2024 to acquire TruEra's AI-observability platform. source
- Agent benchmark
- The homepage reports that Agent GPA caught 267 of 281 human-annotated errors, or 95%, on TRAIL and GAIA traces. source
- Agent metrics
- Built-in agent metrics include tool selection, tool calling, tool quality, plan adherence, plan quality, execution efficiency and logical consistency. source
- API documentation
- The project publishes a searchable API reference covering core classes, metrics, providers and connectors. source
- Batch evaluation
- Runs can replay a dataset and compute the same metrics over stored inputs. source
- Company facts
- LinkedIn lists TruEra as a privately held software-development company founded in 2019 with headquarters in Redwood City, California, and 51–200 employees. source
- Current stewardship
- The official TruLens homepage says the project was created by TruEra and is actively maintained in open source by Snowflake after the acquisition. source
- Custom evaluation
- Developers can define metrics with Python functions, LLM-as-judge prompts, span attributes, rubrics and examples. source
- Custom metrics
- Users can define custom metrics with Python functions or LLM-as-judge prompts and tune judges with rubrics and examples. source
- Dashboard
- The built-in Streamlit dashboard provides leaderboards, trends, individual records and side-by-side version comparison. source
- Database connector
- A Snowflake connector is provided for storing TruLens traces and evaluation data. source
- Evaluation
- It supports agent, RAG, MCP app, and summarization evaluations with metrics including tool selection, groundedness, context relevance, and correctness. source
- Evaluation modes
- Evaluations can run as traces arrive or in batch over a dataset. source
- Experiment comparison
- The leaderboard compares app versions across metrics, latency, and cost. source
- Explained scores
- Its judges attach explanations to metric scores so teams can investigate why a result was rated poorly. source
- Founders
- Snowflake's announcement names Anupam Datta, Shayak Sen and Will Uppington as TruEra's three co-founders. source
- Framework integrations
- Official integrations instrument LangChain, LangGraph, LlamaIndex and NeMo Guardrails applications. source
- Independent company profile
- Craft describes TruEra as an AI-observability company founded in 2019 and headquartered in Redwood City. source
- Installation
- The documented installation uses Conda with Python 3, followed by pip install trulens; source development uses the GitHub repository and editable installation. source
- Integrations
- The documentation lists integrations and providers for LangChain, LangGraph, LlamaIndex, MCP, OpenAI, Anthropic, Amazon Bedrock, HuggingFace, LiteLLM, Google, and Snowflake Cortex. source
- Licence
- The public GitHub repository identifies TruLens as MIT-licensed. source
- Limit: external evaluation costs
- Evaluation can add model-token, infrastructure, database-retention and Snowflake costs, depending on deployment choices. source
- Limit: feature depth and scale
- Conversion Gems characterizes the product as having fewer features than some alternatives, less depth and possible slowness at scale. source
- Limit: operational ownership
- The independent review says production users must manage database lifecycle, access control, retention, upgrades and evaluator-worker reliability. source
Checked against the maker's pages on 15 Oct 2026.