Recoleta
GitHub
Source note

Show HN: PeekAI – Local-first observability for Python AI agents

Authors
ousskh63
Source
Hacker News
Published
Collection
Software Intelligence
AI Agent ObservabilityPython AgentsLocal First DebuggingLLM TracingMulti Agent Workflows
Open GitHubMarkdown source

Summary

PeekAI is a local Python tracing and debugging tool for AI agents. It records LLM calls, tool calls, tokens, costs, errors, and replay runs without sending trace data to a hosted service.

Problem

  • Python agent developers need to inspect LLM calls, tool use, token counts, cost, latency, and failures during development.
  • Hosted tools such as LangSmith and Weights & Biases can require accounts, cloud data upload, and setup before traces are visible.
  • Local storage matters when traces may contain prompts, outputs, tool responses, or user data that a developer does not want to send to a third party.

Approach

  • peekai.init() monkey-patches supported SDK clients at startup, so existing OpenAI, Anthropic, or LiteLLM calls can be traced without changing each call site.
  • Decorators such as @peekai.agent, @peekai.tool, and @peekai.trace create a parent-child span tree for agent workflows.
  • Traces are stored in a local SQLite database at ~/.peekai/peekai.db by default, with an optional custom db_path.
  • The CLI exposes trace listing, trace viewing, stats, map visualization, replay, and cleanup commands.
  • Replay can re-run a past trace, swap the model, or inject a changed tool response, then save the replay as a new trace for side-by-side comparison.

Results

  • The excerpt reports no formal benchmark, accuracy result, latency study, or comparison against LangSmith, Weights & Biases, or OpenTelemetry-based tools.
  • The included multi-agent demo trace records 3 spans: researcher, writer, and format_output.
  • The same demo reports total runtime of 3.6s, 236 tokens, and estimated cost of $0.000222.
  • The demo breaks down LLM usage into 102 tokens and $0.000115 for the researcher call, plus 134 tokens and $0.000107 for the writer call.
  • The local UI runs at http://localhost:8501 and has 4 pages: Dashboard, Traces, Trace View, and Replay.
  • The package lists 3 SDK integration targets in configuration: OpenAI, Anthropic, and LiteLLM.

Link