Ship AI agents you can actually trust.
Sable traces every step your agents take, scores every answer against your evals, and blocks bad actions before they reach a customer. One SDK, every framework, zero guesswork.
Free up to 10k traces a month · No credit card · SOC 2 Type II
support-triage / runs / run_8f2c91
LIVE
Refund request → escalation
gpt-5.1
12 spans
0s
4.8s
agent.run
plan
retrieve.policies
tool.lookup_order
tool.issue_refund
tool.issue_refund
guardrail.pii
respond
issue_refund timed out after 1.9s and was retried. Sable flagged a duplicate-refund risk and paused the run for review.
Powering production agents at teams that ship weekly
Lumen
Fieldwork
Parcelhop
Quanta
Helix
Orbitra
Kiln
Tessera
Driftwood
01 — Observe
See every step. Fix the one that broke.
Agents fail in the middle — a bad retrieval, a flaky tool, a plan that went sideways. Sable gives you the whole run, not just the last message.
User
0ms
Plan
420ms
search_kb
180ms
Draft
1.1s
Traces that read like a story
Every prompt, plan, tool call and retry stitched into one timeline — with inputs, outputs, tokens and cost on every step.
+6.2%
pass rate, last 14 PRs
Evals that run in CI
Score every change against golden datasets before it merges. Regressions fail the build, not your customers.
Redact PII in outputs
Block prompt injection
Pause refunds over $500
Allow web browsing
Guardrails with teeth
Block PII leaks, prompt injection and risky actions in real time — or pause the run and page a human.
prompt v15 · gpt-5.1
0.93
·
You are a support agent for Lumen.
+
Offer refunds only within policy.
·
Check order status first.
+
Cite the policy you applied.
Replay and diff any run
Fork a production run, swap the model or the prompt, and compare outputs side by side before you ship the change.
02 — Integrate
Three lines to full visibility.
Drop in the SDK and every agent framework you already use is instrumented automatically. No proxies, no rewrites.
01
Install the SDK
Python, TypeScript and Go. Under 15ms of overhead per run.
02
Wrap your agent
Auto-instruments OpenAI, Anthropic, LangGraph, the Vercel AI SDK and more.
03
Add evals and guardrails
Start from 40+ built-in scorers or write your own in plain code.
agent.py
agent.ts
agent.go
1
from sable import Sable, eval
2
from agents import support_agent
3
4
sable = Sable(project=“support-triage”)
5
6
# every run is now traced, scored and guarded
7
@sable.trace(evals=[eval.faithfulness, eval.tool_accuracy])
8
def handle(ticket):
9
return support_agent.run(ticket)
10
11
# block risky actions before they happen
12
sable.guardrail(“refund_limit”, max_usd=500, on_fail=“pause”)
pip install sable-sdk
● connected
Works with
spans traced every month
fewer escalations, on average
SDK overhead per agent run
ingest uptime over 12 months
03 — Proof
Teams stopped guessing. Then they shipped faster.
Lumen
We went from “the agent did something weird” to the exact span and prompt in under a minute. Sable is the first thing our on-call opens.
Maya Okafor
Staff Engineer, Lumen
Fieldwork
Evals in CI changed how we work. Prompt changes go through review like code now, and our resolution rate climbed 11 points in a quarter.
Daniel Reyes
Head of AI, Fieldwork
Parcelhop
The refund guardrail paid for the whole contract in its first week. We sleep better knowing risky actions get paused, not shipped.
Tom Lindqvist
CTO, Parcelhop
04 — Questions
Answers, before you ask.
Can’t find what you need? Our engineers answer every message — usually within a few hours.
How is Sable different from logging LLM calls?
Logs show you individual requests. Sable reconstructs the whole agent run — every plan, tool call, retry and handoff — as a single trace, then scores it against your evals so you see why an agent failed, not just that it did.
Which frameworks and models do you support?
Sable auto-instruments OpenAI, Anthropic, Gemini, Mistral, LangGraph, LlamaIndex, CrewAI and the Vercel AI SDK. Anything else works through our OpenTelemetry-compatible SDK in Python, TypeScript or Go.
Will Sable slow down my agents?
No. Spans are buffered and shipped asynchronously, adding under 15ms per run. Guardrails run inline only when you opt in, and typically resolve in under 40ms.
Where is my data stored, and who can see it?
Data is encrypted at rest and in transit, stored in the US or EU region of your choice, and never used to train models. Enterprise plans can self-host Sable inside their own VPC.
Can I run evals in CI?
Yes. Point the Sable CLI at a dataset and an eval suite, and it posts a scored diff to every pull request. Set thresholds to fail the build when quality regresses.
Do you offer a free plan?
The Hobby plan is free forever with 10,000 traces a month, 7-day retention and all core features. No credit card required.
Start in five minutes
Your agents are already in production. Now you can see them.
Free for 10,000 traces a month. Upgrade when your agents do.