- liveOpenTelemetry
- liveGitHub PRs
- liveMCP
- liveREST API
- liveLangfuse
- soonNotifications
- soonLangSmith
- soonHelicone
- Nexprove
- Suture
Your AI is failing quietly.Suture fixes itfor you.
Suture finds the failure in your traces, writes the fix, proves it on your real traces and opens the PR. You never have to read an issue and fix it yourself.
Prefer to do the fixing yourself? Issue-only mode opens a GitHub issue that explains the fix instead.
LLM observability · evals · drift detection · automatic prompt fixes
- Traces checked
- Cost to check
- Merges without you
( illustrative log · hover or focus to pause )
- Suture
- In one answer
What is Suture?
Suture is LLM observability that fixes what it finds. It checks every trace your AI app produces, groups the failures into Issues, writes the fix, proves it on your real traces and opens a pull request, so you never have to fix it by hand.
- Built for
- teams running LLM features in production: agents, chat, writing, support
- Reads
- OpenTelemetry over OTLP/HTTP from any language, or Langfuse with project keys
- Produces
- Issues with a rate and trend, drift alerts, and one proven GitHub PR per Issue
- Fix or issue
- By default Suture fixes it and opens the PR. In issue-only mode it opens a GitHub issue that explains the fix instead
- Keeps you in charge
- nothing merges until you approve, unless you turn on auto-merge
- Costs
- free to try (your first 3 traces, up to 3 fix PRs); paid plans are monthly with daily, weekly and monthly quotas
- Suture
- How it works
Seven steps,
one loop.
From a failing LLM trace to a merged fix, without anyone reading logs or fixing it by hand. Here is one Issue going all the way round.
( trace tr_8f2 )
- source
- spans
- input
- redacted
( one illustrative Issue, start to finish )
- Suture
- Failure graph
See where it
breaks.
Agents, prompts and models, wired to the failures they produce. Thicker lines carry more failing traces. Select a node to see the Issues behind it.
( illustrative · click or tab to a node, enter to select )
- Suture
- Autonomy
Nothing merges until you approve.
Each project has an autonomy level. The default is pr: Suture fixes it and opens the PR, you make the call. Prefer to fix things yourself? Issue-only mode opens a GitHub issue that explains the fix instead. Auto-merge is off until you turn it on, and then it only covers small, prompt-only fixes with checks green.
Writes the fix, proves it on the failing traces and opens a PR. You review and merge.
- Check every trace and group failures: yes
- Write a root cause and fix plan: yes
- Fix in a sandbox, prove it, open a PR: yes
- Open a revert PR if the Issue comes back: yes
- Merge low-risk fixes without review: no
- Suture
- Proof
Every PR arrives proven.
Evals on your real traces, not a guess. A fix is replayed against the traces that failed, three runs each, with tools mocked. The PR opens only if the pass rate goes up and nothing that passed before now fails.
Suture fix: Orchestrator drops the CTA when posts exceed 1,200 chars
| Verification | Before | After |
|---|---|---|
| Failing cases replayed (n=40, 3 runs each) | 25/40 pass | 38/40 pass |
| Existing eval suite[0 new failures] | 112/118 | 113/118 |
| Avg cost per trace | $0.0129 | $0.0131 |
Opened by Suture · autonomy: pr · revert watch: on (7 days)
- Suture
- Integrations
Works with what
you have.
Suture takes the OpenTelemetry you already export, or reads Langfuse with project keys. Nothing to migrate, no Suture SDK.
OpenTelemetry
livePush OTLP/HTTP from any language. GenAI and OpenInference attributes.
GitHub
liveA GitHub App opens one small PR per Issue. No personal tokens.
Claude Code · Cursor
liveMCP server with ten tools: read Issues and fix plans, pull failing cases and patches, start a fix.
Langfuse
livePulls with your Langfuse project keys, no code. Backfills the last 7 days, then checks for new traces every 5 minutes.
Notifications
soonNew Issues, fix PRs and reverts where your team already talks: Slack, Discord, Microsoft Teams, Google Chat and Telegram, plus a signed webhook for anything else. Coming this week.
LangSmith
soonPull connector, after Langfuse.
Helicone
soonPull connector, after LangSmith.
- Suture
- Beyond the loop
Ask it, script it,
teach it.
Tell Suture what each feature should do, ask it questions in plain language, and reach every Issue from your editor or your own scripts.
Features
appSuture reads your PRD and repo docs and drafts checks for each feature. You review and apply them, then it re-checks the last 200 traces per feature from the past 14 days.
Ask
app · mcpAsk in plain language: what broke this week, did the v15 prompt help, where is the money going. Answers come from your own traces and Issues.
MCP server
claude code · cursorTen tools, so your coding agent can take an Issue from diagnosis to an applied patch.
REST API
scripts · ciThe same data at /api/v1, with an OpenAPI 3.1 document. Project tokens are read-only unless you add write.
- Suture
- Where it fits
Keep your tracer.
Add the fix.
Suture is not another place to browse traces. It reads from Langfuse, LangSmith, Braintrust or plain OpenTelemetry and does the work after the dashboard.
| Aspect | Your tracer · Langfuse, LangSmith, Braintrust or plain OpenTelemetry | Suture |
|---|---|---|
| Main job | Record, explore and score traces | Turn failing traces into proven fixes |
| Your traces | Stored and browsed there | Read from OpenTelemetry or Langfuse; a redacted copy kept for replay |
| Checks | Evaluators and datasets you set up | Every trace, against what each feature should do |
| Output | Dashboards, scores and datasets | Issues, fix plans and pull requests |
| The fix | Your team writes it | A sandboxed agent writes it, replays it, and you approve the PR |
Most teams keep both: the tracer to look around, Suture to act on what fails. For the longer answers, see the FAQ.
- Suture
- Pricing
Free to try.
Paid as it fixes more.
Try it free: your first 3 traces analysed end to end, with up to 3 fixes opened as PRs. Paid plans are monthly, with daily, weekly and monthly quotas that refresh.
Free
$0
A real taste of the whole loop, on your own traces.
- Your first 3 traces analysed end to end
- Suture keeps fixing what it finds until up to 3 fixes have opened PRs
- Only fixes that open a PR count. Failed attempts are on us
- Fix by PR, or issue-only if you would rather fix it yourself
- Issues, diagnosis, proof by replay and the 7-day revert watch
Pro
$29
For a team with an AI feature in production.
- Daily
- 500 traces3 fix PRs
- Weekly
- 2,500 traces10 fix PRs
- Monthly
- 10,000 traces30 fix PRs
- Every trace judged, no sampling
- Drift detection and Issues with a rate and a trend
- Fixes opened as PRs, proven on your real traces first
- Features, Ask, the MCP server and the REST API
Team
$99
More projects, more traces, more fixes.
- Daily
- 2,500 traces10 fix PRs
- Weekly
- 12,500 traces40 fix PRs
- Monthly
- 50,000 traces120 fix PRs
- Everything in Pro
- Several projects under one organisation
- Higher daily, weekly and monthly quotas
- Priority on fix runs
Paid plans are monthly. Daily, weekly and monthly quotas each refresh on their own, like a Claude subscription: use up the day and you wait for tomorrow, not for next month.
Need more?
Tell us what you run. A person reads it and writes back, and we raise limits by hand.
- More volumeTalk to us
- Several clientsTalk to us
- Self-host or SSOTalk to us
- ( Free is a one-off taste of the whole loop: it is used once, not refreshed every month. )
- ( Paid plans are monthly. The daily, weekly and monthly quotas each refresh on their own. )
- ( Suture pays for the model calls that check, diagnose and fix. Failed fix attempts never count against you. )
- Suture
- FAQ
Straight answers.
What we store, what we never do, how Suture compares, what happens when a fix is wrong, and what it costs.
- Suture
- Get started
Your next failure gets a PR.
Free to try: your first 3 traces analysed and up to 3 fixes opened as PRs. No card.
- 01Sign up free, then create a project
- 02Send OpenTelemetry from any language
- 03Install the GitHub App, pick an autonomy level
Already have an account? Sign in. Need self-host or SSO? Talk to us.