Probe0 vs Langfuse
Langfuse watches what your instrumented application did. Probe0 sits in the request path on your own machine and changes what your coding agents send, so a call can be answered from a local cache, handed to a model already running in Ollama, or stopped at a spend cap.
Open-source LLM observability and evaluation platform (tracing, evals, prompt management), run as Langfuse Cloud or self-hosted, with data arriving through SDK instrumentation.
| Feature | Probe0 | Langfuse | Edge |
|---|---|---|---|
| Where your prompts go | Stay on your Mac. No Probe0 server exists. | Langfuse Cloud, or self-hosted Postgres + ClickHouse + Redis + S3. | Probe0 |
| Setup for one developer | One proxy plus one certificate, installed once. | SDK instrumentation, decorators, or OpenTelemetry spans. | Probe0 |
| Covers Claude Code, Codex, Cursor | All of them, no per-tool config. | Closed CLIs emit no spans, so nothing to instrument. | Probe0 |
| Response caching | Exact cache on local disk; hits never touch the network. | No LLM response caching. Docs point at a gateway for it. | Probe0 |
| Tool-call safety in semantic cache | Strict similarity floor; refuses anything carrying tool calls. | No semantic cache to make safe. | Probe0 |
| Routing and failover | Local routing, model tiering, auto-retry on the cloud. | Not in the request path. Integrates with gateways instead. | Probe0 |
| Spend limits | Hard cap per run and per day. Warns, then pauses. | Cost tracking, dashboards and alerts. No request-path cap. | Probe0 |
| Cost attribution and plan advice | Per-process ledger; flags when your tier exceeds your usage. | Cost per trace, session and user in instrumented code. | Probe0 |
| Trace depth inside an app run | One row per HTTP call: model, tokens, cost, latency, process. | Nested spans across tool calls, retrieval, sub-agents, non-LLM work. | Competitor |
Langfuse builds the best open-source trace tree I have used, with nested spans across retrieval and tool calls, datasets, LLM-as-a-judge scoring and versioned prompts, and I would rather say so than pretend otherwise. What that costs you is a request path Langfuse never touches. If your bill comes from coding agents rather than an application you ship, Probe0 is the tool that changes the number: the proxy answers repeats from a local disk cache, sends work to a model already loaded in Ollama, and stops a run at a cap you set. Nothing leaves the machine to make that happen, because there is no Probe0 server to leave for, and setup is one certificate rather than instrumentation spread through code you may not own.
None of the observability work reduces a bill. Langfuse does not cache responses, does not route or fail over, and does not enforce a spend cap, because it is not in the request path. Its own writing is straightforward about this: enforcement belongs in a gateway, and observability tells you what number to set the cap to. So the loop with Langfuse alone is that you instrument, you watch the cost dashboard climb, and then you go somewhere else to do something about it. There is also the capture problem. Langfuse sees what you instrumented. Claude Code, Codex and Cursor are closed CLIs running on your machine, so unless you wrap them there is nothing to decorate, and their traffic simply does not land in your traces.
Probe0 takes the other half. It is a local LLM proxy on macOS, currently in private beta: one process plus a certificate installed once, and every coding agent CLI on the machine routes through it. From there each module is switchable on its own and reports what it saved. Exact cache answers repeats from local disk without touching the network. Semantic cache uses a local vector index with a strict similarity floor and refuses to match anything carrying tool calls, because a wrong hit corrupts a run and a corrupted run costs more than the call you skipped. Local routing sends work to a model already loaded in Ollama or LM Studio and retries weak answers upstream. Model tiering tries the cheap model first. Coalescing collapses simultaneous identical calls into one. Spend Guard caps a run or a day, warns, then pauses. Recording keeps a full ledger of model, tokens, real cost, latency and which process made each call, which is also what lets it tell you when your plan tier is above what you actually use. It is macOS only, it is one machine and one user, it signs in with Google or GitHub on a young account system, and its provider coverage is what coding agents call rather than a hundred-provider catalog.
The verdict
Pick Probe0 if you are a developer on a Mac whose bill comes from AI coding agents. It is the only one of the two that can actually spend less on your behalf: caching, local routing and a cap that pauses, all in the request path, with a per-process ledger of what every agent cost you and no data leaving the machine. Langfuse is the better pick in one situation, and it is a real one: a team shipping an LLM application that needs nested traces, evals, prompt versioning and shared history across environments. That is a different job from the one in front of you. If the problem is what your coding agents are costing you today, install Probe0.
Questions
- Is Probe0 a Langfuse alternative?
- Only for cost control and for seeing what your coding agents actually spent. Probe0 is an LLM proxy that caches, routes and caps spend in the request path, while Langfuse is an observability and evaluation backend for instrumented applications. If you need datasets, evals, prompt management or team dashboards, Probe0 is not a substitute.
- Can Langfuse cache LLM responses or cap my spend?
- No. Langfuse tracks tokens and cost per trace and can alert you, but it does not cache model responses and does not stop a request. Its documentation treats caching and enforcement as gateway or proxy work. Probe0 does both locally, with an exact disk cache, a conservative semantic cache and a hard per-run and per-day cap.
- Does either tool track Claude Code, Codex or Cursor usage?
- Probe0 does by default, because every agent CLI on the machine routes through the local proxy and lands in the ledger with model, tokens, real cost, latency and calling process. Langfuse would need those CLIs to emit instrumentation or OpenTelemetry spans, which closed coding agents generally do not give you.
- Can I run both, and does either work fully offline?
- Yes to both, in different ways. Self-hosted Langfuse pointed at a local Ollama endpoint is a zero-egress stack, and Probe0 runs entirely on your machine with no Probe0 server at all. Running Probe0 for local LLM routing and agent cost control alongside Langfuse for your product's tracing and evals is a reasonable setup.