Deployment models for AI agent observability

There are several tools in the AI-agent observability and governance space. They differ along feature axes, and they also differ along a more fundamental axis: where the tool actually runs, and which party sees the data. This page describes the latter. It is not a feature comparison.

Where each tool runs, and what its vendor observes

Tool	Runtime location	Data observed by vendor	Account required	Works offline
LangSmith	Vendor cloud	Prompts, responses, tool calls, traces	Yes	No
Helicone	Vendor cloud proxy (managed); self-host available	Cloud tier observes all prompts and responses passing through the proxy. Self-host is local.	Yes for cloud	Self-host only
Langfuse Cloud	Vendor cloud	Prompts, responses, traces, evaluations	Yes	No
Langfuse self-hosted	Your infrastructure	Only the operator of the deployment	No vendor account needed	Yes
Honeyhive	Vendor cloud	Prompts, responses, traces	Yes	No
Arize Phoenix self-hosted	Your infrastructure	Only the operator of the deployment	No vendor account needed	Yes
Occasio	Local process on your machine	Nothing reaches Occasio Labs or any third party. Outbound traffic goes to whichever LLM endpoint you configured, authenticated with your own API key.	No	Yes

The cell values above describe where each tool was designed to run and which party can observe traffic in its default configuration. For tools that ship both a cloud tier and a self-host option, the rows separate the two so the architectural placement is unambiguous.

What this table is and is not saying

This is a descriptive comparison along one dimension: where the tool runs and what its operator sees. SaaS observability tools are a legitimate category. They exist for valid reasons (zero-setup onboarding, cross-team aggregation, hosted dashboards, managed retention), and a team that has chosen the SaaS model will find them well-engineered for that model.

This table is meant to help a reader who is matching a requirement to a category. Three common requirements that lead to the local-first row:

• The data being processed is regulated or contractually restricted from leaving the local environment (healthcare, defence, EU data residency).
• The team wants no third-party dependency in the path between an AI agent and the LLM provider.
• The team wants the audit artefact (the thing presented to an auditor or compliance reviewer) to be self-contained and verifiable without a vendor relationship.

If none of those apply, a SaaS observability tool may serve the team better than Occasio. If any of them apply, the architectural placement matters and a local-first tool is the only category that fits structurally.

Related

• docs/WHY-LOCAL.md describes the Occasio architecture in detail.
• docs/SUSTAINABILITY.md describes how a local-first product is funded.