TrueClara layer
reward layer
entry/home
route/pricing
value/checkout
edge/confirm
api/api/sub
goal/success
The production reward signal for coding agents.
TrueClara runs every agent-shipped change as a randomized canary, measures the real production outcome, and returns a calibrated verdict — keep, roll back, iterate, or abstain — over an API and MCP. Causal by randomization. Honest when it can’t answer.
Start free →
Explore the live demo
always-valid sequential stats · abstains when underpowered · cookieless
the route graph
Your app is a graph. Every deploy is the trigger.
TrueClara parses your Next.js app into a route graph, then overlays each deploy on top of it — so the experiment isn’t scoped to a list of commits, it’s scoped to the routes and edges the change actually touched. Ship the change and the canary starts itself.
changed by deploy
verdict: rollback
★ value route
verdict: keep
new route this deploy
what it proves
A randomized canary, an honest verdict.
CI checks the build. Error monitors catch what throws. Neither tells the agent whether the change actually helped in production. TrueClara runs every agent-shipped change as a canary against control, measures the real outcome, and returns a verdict it can trust — including when the honest answer is “not enough signal yet.”
Trial → paid conversion, treatment vs. control
randomized canary/pricing → /checkout
Runninglift −14%
Value-route reach, treatment vs. control on /onboarding
randomized canary/signup → /onboarding
Runninglift +9%
step 01
Ship as a randomized canary
The deploy is the trigger: an agent-shipped change splits live traffic into treatment and control automatically, so any measured difference is caused by the change — not inferred from uncontrolled traffic.
see how canaries are scoped →control 31%treatment 17%1,840 sessions
step 02
A calibrated verdict, not a guess
Always-valid sequential statistics watch the treatment/control split and call it the moment the data supports an answer — keep, rollback, or iterate. If the data can’t support a defensible answer yet, TrueClara abstains instead of guessing. Release Radar is where every verdict lands — each release scored by what it did to the outcome, not whether the build passed.
read the reward API →CPredicted impact · pricing v2
/pricing→
/checkout→
/success
This PR reweights the /pricing → /checkout edge. 1 high-traffic value route predicted affected. Merging starts the canary.
step 03
Predicted impact, before you merge
TrueClara posts predicted route impact on the pull request before the canary even starts — then runs a retrospective comparing the prediction against the verdict the canary actually returned.
see PR-impact prediction →what it replaces
Today the agent ships blind. No reward signal exists.
Release data, route topology, runtime behavior, and outcome data each sit in a different tool — and the work of turning them into “did this change actually help” is glue code nobody owns, so agents optimize against proxies instead: tests, lint, vibes. TrueClara is that missing reward signal. Keep the tools your team already uses; replace the guessing with a verdict.
release
Deploy & PR context
topology
Next.js route graph
runtime
Web & mobile events
outcome
Production outcome signals
delivery
Slack, webhooks, email
Trial → paid conversion, treatment vs. control after deploy 7f3a2c1.
−14%
measured lift, 92% confidence
Experiment/pricing → /checkout
Triggered bydeploy 7f3a2c1 · #482
Actionrollback target ready
One command to install. A reward signal your agent can call.
Drop the SDK into any Next.js app on 13.4 or later — App Router or Pages. Cookieless by default. Then connect a deploy and point your agent at the reward API or MCP server; the three steps below take a few minutes, mostly one PR.
$
npx @trueclara/install
1run the installer
2upload your route graph + connect a deploy
3ship — the canary starts and Clara returns the verdict
always-valid
Sequential statistics (mSPRT). Check the verdict any time without inflating false positives.
abstain
The verdict when the data can’t support a defensible answer — honest, not a confident guess.
∞
Every measured experiment joins the corpus. The corpus compounds — each verdict makes the next one faster.
cookieless aggregate mode by default · no read access to your source or production database — the SDK reports out, nothing reaches in. It flags broken routes and dead ends along the way — table stakes, not the headline.
Clara is scoring the next release
Ship fast. Know within the day what actually worked.
Install the SDK, link your repo, connect the reward API or MCP — and give your agent a production reward signal instead of a hunch.