Agent surface & MCP

langprobe's surface is built for agents, not just people. A 48k-token trace is useless to an LLM with a budget; langprobe projects it into a token-budgeted, LLM-legible slice so a coding agent can find a failed run, read the salient part, replay an edit, and read the diff — hands-free.

Token-budgeted run views

Every agent read is a projection over the same data the UI shows, sized to a token budget instead of dumped raw:

• GET /v1/runs/{run_id}/agent-view — a salient slice of a run: the spans that matter (errors, the critical path), trimmed to fit a budget. A 48k-token trace becomes a ~2k-token summary an agent can actually reason over.
• GET /v1/agent/failed-runs — the failed runs worth looking at, ranked, so an agent can pick what to debug.
• GET /v1/agent/instrument-guide — a machine-readable guide for wiring up tracing, so an agent can instrument a repo without a human in the loop.

These are the same REST endpoints humans and CI use — one surface, three audiences.

MCP: an agent can debug an agent

Over MCP, langprobe exposes the debug loop as tools a coding agent (Claude, etc.) can call directly:

• runs.search — find the run (e.g. status error, last hour).
• runs.read — read its token-budgeted salient slice.
• replay.edit — apply an edit (change a timeout, a prompt, a model) and queue a replay.
• replay.diff — read what changed, with the determinism verdict.

That's the whole loop — find → read → replay → diff — as API calls, so the agent that's debugging never has to leave its context to open a dashboard.