PetalTrace

Overview

PetalTrace is an open-source observability tool for developers tracing AI agent workflows. It suits teams that want local control and replay-oriented debugging; its biggest deployment caveat is that authentication is future functionality.

It brings trace inspection, comparison, and replay together across a CLI, HTTP API, MCP server, and web interface. That breadth makes it useful in both developer workflows and agent-driven analysis, while its SQLite-backed local store favors self-managed deployments over a centrally managed service.

Key features

PetalTrace captures prompts and completions, tool calls, token use, costs, and execution timelines. Full prompt capture includes system instructions, message history, and tool definitions, giving developers context for diagnosing model behavior rather than just inspecting outputs. PetalFlow users can choose minimal capture for latency, status, and token counts; standard capture adds prompts, completions, and tool I/O; full capture adds graph snapshots and edge data. The choice is practical: lighter capture limits recorded content, while full capture preserves more workflow structure.

Run comparison surfaces prompt, output, structural, and cost differences between executions. Replay can use another model or temperature and supports live, mocked, and hybrid modes, making it possible to investigate changes under different execution conditions. Cost tracking by workflow, provider, or model, full-text search across prompts and completions, and live SSE feeds round out the debugging toolkit.

The React web UI explores traces, costs, and workflow graphs; the CLI and HTTP API provide other ways to work with the same capabilities. Its MCP server lets agents query trace history, inspect prompts, analyze costs, compare runs, and initiate replays, with documented Claude Code configuration. PetalTrace also accepts standard OTLP traces from OpenTelemetry-instrumented applications, including those that do not use PetalFlow, which widens its usefulness beyond that framework.

Runs, spans, and LLM interactions are stored in SQLite, with the default database at ~/.petaltrace/data.db. Retention defaults are 30 days for runs and 90 days for failed runs, with a maximum configurable period of 365 days. Local storage offers direct control over trace data, but teams must manage the service and its data themselves. With authentication disabled by default, it is a poor fit for deployments that need access controls unless they can address that limitation outside the product.

Pricing

PetalTrace is free and open source under the MIT license. Its public plan costs nothing, and no commercial price or usage cap is stated. That makes it an appealing starting point for developers who can operate their own instance; the tradeoff is self-management rather than a priced hosted tier with service terms. The repository documents building from source with Go or downloading a release binary and running the daemon locally.

Platforms

PetalTrace supports API, Linux, macOS, Windows, web, and self-hosted use. The documented local daemon and React web UI suit developers who want a service they control, while the API, CLI, and MCP server accommodate programmatic access and agent workflows.

Who it's for

PetalTrace is best for developers diagnosing agent behavior across prompts, tools, costs, and workflow execution, especially those using PetalFlow or OpenTelemetry and comfortable with self-hosting. Its MCP interface is also a strong fit for teams that want agents to inspect and compare trace history. It is less suitable for organizations that require built-in authentication or do not want to operate local infrastructure.

Pros and cons

• Pros: Captures prompt, tool, token, cost, and timeline data, with selectable PetalFlow capture depth to balance context against capture scope.
• Pros: Run comparison and replay across models, temperatures, and live, mocked, or hybrid modes support iterative debugging.
• Pros: CLI, HTTP API, MCP, and web UI cover distinct working styles; OTLP support extends tracing to OpenTelemetry applications beyond PetalFlow.
• Pros: Free, MIT-licensed, and locally stored in SQLite, giving self-hosting developers direct control without a stated usage cap.
• Cons: Authentication is disabled by default and described as future functionality, making access control a deployment concern.
• Cons: Local deployment and SQLite storage put operation, data management, and retention decisions on the user.

Alternatives

AI Agent Observability Tools is the broader category directory for comparing options. OpenLIT is another free, self-hosted option; its OSS plan explicitly includes unlimited users, projects, and environments, which suits readers who need those stated allowances.

W&B Weave may suit users who want evaluations and scorers alongside tracing, with its free tier capped at 1 GB monthly ingestion and 5 GB storage. SigNoz is worth considering when a broader self-managed observability setup is preferable, though its community plan puts infrastructure, storage, scaling, upgrades, and backups on the user.

Opik is an alternative for developers seeking open-source observability and evaluation features they can download and run locally. Agenta offers a Hobby plan with 2 team members, 5,000 agent runs per month, 20 evaluations per month, and one-week trace retention; it is a more bounded choice for a small team that wants explicit limits.

Langfuse offers a Hobby tier with 50k units per month, 30 days of data access, and 2 users, making its stated limits useful for teams sizing a hosted or self-hosted alternative. Traccia gives hobby users 50K events and seven days of retention, while its Observe plan is $99 a month for 500K events; consider it when event quotas and retention terms are central to the choice.

TraceRoot is another option, with a Starter plan at $30 a month.

Verdict

Choose PetalTrace if you are a developer who wants a free, self-hosted way to inspect and replay AI agent runs, especially through PetalFlow, OpenTelemetry, or MCP workflows. Its combination of detailed trace capture, comparison, and replay is the core reason to choose it. Look elsewhere if built-in authentication or a managed service is essential; those needs outweigh the advantages of its local-first design.