Halo
- HobbyAvailable
- CoreAvailable
- ProAvailable
- EnterpriseAvailable
- Self HostedPublic Beta
Self-hosted: Halo is in public beta from version >=v4.28.0 and
needs an instance-wide Langfuse AI model and a worker queue consumer. See
self-hosted setup.
Halo is the Langfuse assistant for exploring your project data and taking actions in Langfuse with your approval. You can ask about traces, observations, sessions, and metrics, or ask it to create datasets, dashboards, and other resources. It brings the same Langfuse project tools available to external coding agents into the Langfuse UI.
Product intro with music and no narration. On the Tracing page, Halo opens as a side panel with suggestions such as “Analyze failure patterns”, “Review recent activity” and “Investigate unusual patterns”. Example prompts follow: “Look at the traces of the last 7 days. Anything interesting?”, “Set up an evaluator to capture user frustration.”, “Create a dashboard for me” and “Help me optimize the cost of my agent”. The closing screen shows that ⌘ I opens Halo.
What you can do
For work inside Langfuse, use Halo for the same investigation and resource-management tasks you would delegate to an external agent:
| Goal | Example prompt |
|---|---|
| Investigate your data | "Which traces had the highest latency yesterday?" |
| Analyze errors at scale | "Fetch failed observations from the last 7 days and cluster the main errors." |
| Build a resource | "Create a dataset from last week's failed generations." |
| Monitor performance | "Build a dashboard of error rate and p95 latency for the last 7 days." |
| Navigate Langfuse | "Open the traces table filtered to today's errors." |
| Learn how Langfuse works | "How do I set up an evaluator?" |
Optimize costs
Halo breaks down the last 30 days of spend by workflow and returns a table of savings opportunities with estimated savings, effort, and quality risk, plus a suggested order.
Create a dashboard
Halo reads the project's traces and scores, asks once for approval to create widgets, and builds a dashboard with cost, conversations, quality signals, latency, errors by step, and tool usage.
Get started with evaluations
Halo asks two setup questions, checks existing scores and observations, and recommends quality signals to track, split into Langfuse evaluators and application code.
Investigate and fix a regression
Halo traces rising user disagreement to a stale production prompt, tests a newer prompt version against the disputed traces, and promotes it to production after approval.
How it works
Halo uses the same Langfuse MCP server project tools that external agents can connect to, together with documentation search and sandboxed code execution. The in-app experience adds your current UI context (the trace, observation, session, or dashboard you are viewing, plus your name, browser language, and timezone), background execution, notifications, and approvals in Langfuse.
It can:
- Query traces, observations, sessions, and metrics.
- Process large observation sets with code in a sandbox.
- Create or update datasets, dashboards, widgets, prompts, and score configs.
- Link you directly to filtered views and relevant pages in Langfuse.
Action approvals
Halo pauses for your approval before changing data or configuration. Always approve lets a tool run without asking again in the current conversation. Halo is limited to your existing permissions in the current project. Approval does not grant additional access.
Background runs
Runs continue independently of your browser, so you can keep using Langfuse or start another conversation.
Conversation history
Halo conversations are saved per user and project so you can return to previous conversations.
Conversations follow the project's data retention policy, measured from each conversation's last update. When an expired conversation is deleted, its runs and events are deleted too. Cleanup waits for unfinished runs and credential cleanup, so deletion is not immediate at expiration. Disabling Halo does not pause retention.
Data privacy and security
Halo follows the same data handling model as other Langfuse AI features:
- On Langfuse Cloud, model requests use Langfuse-managed Amazon Bedrock in your Langfuse data region. On self-hosted deployments, requests use the instance Langfuse AI model you configure.
- Access is scoped to the signed-in user and current project.
- AI feature tracing for product and service improvement can be enabled or disabled separately by organization admins and owners.
Limitations
Halo can misunderstand a question or return incomplete results. Verify important answers and review proposed actions before approving them.
Feedback
You can submit thumbs-up or thumbs-down feedback on Halo responses. If AI feature tracing is enabled, feedback may be used for product and service improvement. You can also share your thoughts in our feedback discussion.
Related resources
- Self-hosted Halo setup
- AI Feature Security
- Langfuse MCP Server
- Langfuse Agent Skill
- Halo announcement
- Share Halo feedback
Last updated on