failover added 1.9s. Full trace: tr_8f2a...
Debug production traffic
Ask why latency spiked. Get the exact trace back, not a dashboard safari.
Inference MCP plugs your workspace into any LLM client. Ask about projects, inferences, traces, evals, and training jobs in plain language, from the coding agent you already use. OAuth sign-in in the browser. No API key.
Ask why latency spiked. Get the exact trace back, not a dashboard safari.
Scores dropped overnight? Ask which samples regressed and read the diffs without leaving your editor.
Poll your fine-tune from the same session that launched it. Loss curves as text beat tab switching.
Spend, tokens, cache hits, latency percentiles. Ask for the figure instead of exporting a CSV.
One command or one click. Works in any MCP client that speaks HTTP.
OAuth in the browser. Your session gets your team permissions, nothing more. Revoke anytime under Settings → Connected Agents.
Your agent picks the right tool per question. You read answers, not dashboards.
Every project you can see, with traffic summaries.
Filter request logs by model, status, latency, or time.
Full span tree for one request, timings included.
Eval runs with scores, deltas, and regressed samples.
Status, steps, loss, and artifacts for a fine-tune.
Your Halo agents and their recent runs.
Alerts and anomalies your project has raised.
Spend and token counts by model, project, or day.
Datasets built from captured traffic, ready to train on.
Reads by default. Write scopes are requested at sign-in and follow your team role. A viewer stays a viewer.
Any MCP client that speaks streamable HTTP. The six above are tested. Everything else takes the raw URL.
Nothing. The MCP server is free. You pay only for inference you run on the platform.
Settings → Connected Agents lists every connected client. Revoke one and it is signed out on its next call.
Two minutes. OAuth only. Uninstalls as fast as it installs.