Agent Access

One MCP endpoint over the whole lake.

CtrlB does not sell you an agent. It exposes a Model Context Protocol interface over the same data layer the UI queries, so the agent you already trust reads complete operational history directly — no dashboard scraping and no export pipeline to keep in sync.

Let your agents investigate without limits.

Traditional platforms sample traces, cap cardinality, age out logs, and throttle queries — walls that keep a vendor’s bill flat at human query volumes. An agent investigates at machine scale and hits every one on the first incident. CtrlB has none of them: reads are unmetered, history stays whole, and the only limit left is the evidence itself.

Traditional observabilityLive trace
Why did checkout latency increase?
00:01Querying checkout p99 by region
00:02!Trace sample — 88% dropped by collector
00:03×429 Too Many Requests
00:03Retrying in 2s
00:05Comparing three-month baseline
00:08!No data — beyond 14-day retention
00:11Segmenting by customer_id
00:14!customer_id dropped — cardinality cap
Investigation degraded
Queries completed
3
Dimensions explored
2
History compared
14d cap
Rate-limited
1
Elapsed
00:19
Outcome

The wall isn’t a bug — it’s how the vendor keeps its own bill flat.

CtrlB over MCPLive trace
Why did checkout latency increase?
00:01Comparing latency across regions
00:01Comparing current vs healthy deploys
00:02Segmenting by customer_id
00:02Inspecting retry behavior
00:03Correlating Kubernetes events
00:03Comparing 90-day baseline
00:04Searching similar incidents
00:04Synthesizing evidence
Investigation complete
Queries completed
3,442
Dimensions explored
36
History compared
90 days
Rate-limited
0
Elapsed
00:04
Root cause

payments-api v84 introduced a retry storm after deployment.

Connect the agent you already use

CtrlB speaks native Model Context Protocol, so any MCP client reads the same lake the UI queries — no bundled assistant, no lock-in.

Claude CodeClaude Code
CursorCursor
VS CodeVS Code
WindsurfWindsurf
+ any MCP client

Native MCP support

Structured access to the whole lake through Model Context Protocol. Works with Claude Code, Cursor, and any MCP client.

Docs: MCP

Same lake as the UI

Agents and engineers read the same open Parquet. No dashboard scraping, no export pipeline to keep in sync.

Data Engine

Built for AI workloads

Sub-second queries and unmetered reads so agents can iterate freely — twenty questions instead of three.

Playground

What it gives every agent

  • Bring your own agent; the interface is open rather than a bundled assistant.
  • Sub-second responses, so an agent can ask twenty questions instead of three.
  • Agents and engineers read the same evidence from the same files.
  • Sustained query volume is a compute decision, not a licensing tier.

What you can build

Real workflows when agents can query freely — without scraping dashboards or standing up a second copy of the data.

your agent · ctrlb mcp
why did checkout p99 spike after the 09:12 deploy?
─ ctrlb.search_logs (checkout · 24h)
─ ctrlb.get_trace (a91b…7d40)
payments-api v84 added a retry storm at 09:12 — p99 +340ms, isolated to us-east.
3 tool calls1.6s$0.00 reads

Natural-language investigation

“Why did checkout latency spike after the 09:12 deploy?” The agent translates intent into queries, walks the evidence, and returns a root cause — no query language required.

Alert · error-rate > 2% · checkout
Auto-triage · 96% confidence
  • correlates with payments-api v84 rollout
  • contained to us-east, already recovering
  • matches incident #4821 signature

Alert triage automation

The agent reads an alert, pulls related history, checks past incidents for context, and drafts a verdict — escalating real threats, suppressing the noise, before a human opens the UI.

ctrlb · one query, three signals
SELECT l.msg, m.p99, t.duration
FROM logs l
JOIN metrics m USING (ts, service)
JOIN traces t USING (trace_id)
WHERE service = 'checkout'
1 correlated event3 signals1.4s

Cross-signal correlation

One endpoint over logs, metrics, and traces. Joins that used to need three tools and three export jobs are a single query against one lake.

Claude CodeCursorCI pipelineon-call botCtrlB MCP

CI and on-call bots

Wire Claude Code, Cursor, or your in-house bot to the same MCP surface — investigations that run where engineers already work, at the same speed as the UI.

ask over patterns, not raw lines
GET /api/checkout %d %dms×1.9M
cache miss key=%s region=%s×412K
retry svc=payments attempt=%d×6.2K

Pattern-aware questions

Ask over compressed log patterns and full history instead of a sampled fifteen-day window. The agent reasons over the shape of production, then pulls raw evidence only where it matters.

notebook · ctrlb mcp
In [1]:ctrlb.query("p99 by region", window="90d")
Out[1]:DataFrame · 12 rows · 480 ms

Custom workflows

Notebooks, scripts, and internal agents hit the same MCP surface as the UI. Query programmatically, enrich alerts, or wire the lake into processes you already run.

Other core components

Back to platform overview

Give your agents the lake they deserve.

Connect Claude, Cursor, or your own bot. One MCP endpoint. Complete history. Questions that stay free to ask.