Tracing

Every answer leaves
a paper trail.

Each request becomes an OpenTelemetry trace — auth, retrieval, tool calls, model, tokens, latency and cost. Nothing about a reply stays a mystery.

14 days free · No credit card · Bring your own model keys

Inside one request

Read a request span by span. In the order it ran.

A trace is not a log line. It is the sequence of steps the agent actually took, each with a duration you can compare against the others.

Auth

Which key, which user, which scope — and the milliseconds spent checking all three.

auth.verify
keysk-so-•••
useralice@acme.com
scopechat:write
duration12 ms

Retrieval

The documents and FAQs that were pulled, their similarity scores, and the page each passage came from.

knowledge.retrieve
policy.pdf · p.120.94
FAQ#refund0.81
manual.pdf · p.40.73

Tool calls

Every MCP call with its arguments, its response and its latency. The one that timed out is obvious.

Tool calls
search_orders("1234")
87 ms · 200
get_user(42)
31 ms · 200

Model

Prompt tokens, completion tokens, time spent streaming and the cost of the call. Per request, not per month.

llm.stream
prompt
412
completion
87
latency
423 ms
cost
$0.012

Across all of them

Find the slow one. Then fix the cause.

Narrow thousands of requests down to the handful worth reading, attribute the spend, and take the spans wherever your team already looks.

Query the history

Filter by agent, user, status or latency. “Everything over two seconds yesterday” is one query, not an afternoon.

latency > 2s
req_8a2c3.2 s
req_9b412.7 s
req_a1f02.1 s

Cost per agent

Token spend split by agent and by day, so the bill arrives with a name already attached to it.

Token spend · this month

$84.21

Grouped failures

Errors cluster by type and by tool, so you repair the integration that is actually breaking rather than the loudest one.

Failures today
mcp.timeout×4
rate_limit×2
tool.500×1

Export the spans

OpenTelemetry out to the APM or warehouse you already run. Your traces, your retention, your dashboards.

OTLP export
Datadog
Grafana Tempo
BigQuery

What teams say

Every time you want to add AI to your app, you end up writing a ton of code — managing API keys, handling conversation history, engineering prompts, counting tokens. And then six months later a better model comes out and you have to refactor everything. SentientOne just removes all of that.
Daniel Marsh · CTO

Pricing

Plans for every team size.
Fourteen-day trial, cancel anytime.

Starter

$19/month

For developers putting the first AI feature into a product.

Start free trial

1 agent · 10k requests/mo

  • 1 MCP server per agent
  • Knowledge — 100 credits †
  • Embeddable chat widget
  • Analytics · 30 requests/min
  • 14-day logs & chat history
  • Community support
  • Streamable HTTP endpoint
  • AI Workspace (team chat)

Pro

Most popular
$49/month

For teams running AI in production and watching the numbers.

Start free trial

5 agents · 50k requests/mo

  • 2 MCP servers per agent
  • Knowledge — 300 credits †
  • Full tracing, cost & token analytics
  • Streamable HTTP endpoint
  • AI Workspace (team chat)
  • Extra seats at $9/member
  • 90-day logs · 50 requests/min
  • Email support, 48h response

Enterprise

Custom

For organisations with compliance, scale or data-residency rules.

Talk to us

Unlimited agents & requests

  • Unlimited MCP servers & knowledge
  • Self-hosted / single-tenant option
  • Custom rate limits & log retention
  • SSO and organisational support
  • Security review assistance
  • Dedicated SLA & onboarding

Prices in USD. Every plan includes bring-your-own model keys, the agent chat interface, and a 14-day trial. † 1 credit = 1 FAQ · 5 credits = 1 document.

Get started

Stop guessing. Read the trace.

Tracing is on by default on every plan. Send one message and the whole span tree is waiting in the dashboard when you get there.

  • Create an agent
  • Send one message
  • Open the trace
  • Export the spans

14 days free · No credit card required