Get the weekly radar
One short email with the freshest products, every week.
AgentLens records AI agent runs locally so you can replay and debug them offline, without burning API credits.
Grounded in available product and source data
Debugging an AI agent usually means re-running it against the live API every time something looks wrong, burning credits on repeated calls just to watch the same failure happen again; AgentLens is built to break that loop by recording a run locally the first time, then letting a developer replay and step through it afterward without hitting the API again.
What actually gets captured is specific: inputs, outputs, tool calls, and token usage, stored locally rather than sent to a third-party debugging service. On top of that recording layer sits cost tracking, which estimates API spend from token counts against configurable pricing models — useful for spotting an expensive tool-call loop before it shows up as a surprise on a bill, not just after the fact.
Integration currently covers LangChain's LLMs, Chains, and Agents, plus the OpenAI Python client directly, with plain Python function decorators available for instrumenting custom code that doesn't go through either. The README states CrewAI and other framework integrations are planned but not yet shipped, so a team building on a framework outside that current list would need to wait or write their own instrumentation.
As an open-source project (MIT-licensed, installable via pip as 'auriel-agentlens'), it's worth being clear-eyed about where it actually stands: the repository currently shows a small star count and is marked pre-release, meaning this is an early, actively developing tool rather than something with a long production track record. Anyone evaluating it for a team's debugging workflow should weigh that maturity level against the value of local-first replay that doesn't consume additional API credits, before committing to it over a more established alternative.
No — the tool is described as local-first, recording and replaying agent runs on local storage rather than uploading run data to an external debugging platform.
Not yet — the README states CrewAI and other framework integrations are planned but not currently available, so today's support is limited to LangChain, the OpenAI Python client, and custom Python decorators.
Yes — the README describes plain Python function decorators for instrumenting custom code directly, so recording isn't limited to the two named framework integrations.
It estimates API usage cost from token counts run against configurable pricing models, alongside the raw token-usage numbers themselves, rather than pulling real billing data directly from a provider account.
Via pip, as the package 'auriel-agentlens,' available on PyPI — no separate signup or hosted account is described as required to start using it.
