Prerequisites
- An agent whose code you can edit and run. The project opens only once that code sends a span tagged with a call site.
- For the recommended instrumentation path, a coding agent that can read the repository, edit files, and fetch a URL.
Connect your first agent
1
Sign up
Go to app.tessary.ai and create an account. Tessary creates an organization and a first project for you.
The project opens on the connect screen, Connect your traces. It stands in front of the whole product until a tagged span arrives.
2
Copy the endpoint and create a token
The connect screen carries both values your exporter needs:
- Endpoint is
https://app.tessary.ai/v1/traces. - Under Bearer Token, select Create and copy. This mints a project-scoped API key with write scope and copies it. Treat it as a secret: never hardcode it in application source and never commit it.
3
Instrument your agent
Instrumenting means making two things true: traces reach the endpoint, and every span that covers a model call carries a
tessary.call_site.id attribute. There is no Tessary client library to install.The recommended path is to copy the prompt under Paste into your coding agent and hand it to a coding agent with your repository open. Instrument your agent covers that path and wiring the exporter by hand.4
Run your application
Exercise the instrumented paths. Your test suite is usually enough.If it reads Spans are arriving without a call site instead, your traces are reaching Tessary but no span carries the tag. Instrumentation troubleshooting works through that state and the case where nothing arrives at all.
The screen stops reading “Listening on /v1/traces. Nothing has arrived yet.” and opens on its own when the first tagged span lands.
What happens after the first trace
Triage names the stage your project is in whenever its queue is empty. Once traces are arriving it readsfitting: classifiers are learning what your traffic normally looks like, so they can tell a change from ordinary variation. That takes a while on low volume, and it is expected rather than a fault.
Six classifiers run without setup: duration_drift, cost_drift, tool_error, secret_leak, malformed_output, and groundedness. A seventh, frustration, starts switched off and needs your own OpenRouter or TypeSafe key.
When a classifier detects a change, it creates a finding on the Classifiers page. A finding records what changed and the evidence for it. It is not yet a validated issue.
Triage a finding and investigate the case
On Tessary Cloud, triage starts when you start it. Automatic triage is not available, so open a finding and select Run triage. Triage reviews the finding’s evidence and rules whether it is a real issue. A positive ruling opens a case, or joins the case already open for that cause. A high-confidence secret leak and a frustration finding skip triage. Each opens its case as soon as it is filed, because its numbers are the claim. On a case, select Run RCA to start the investigation. The report gives a verdict, the hypotheses it considered with the traces it cites for each, and the checks it ruled out. Connect a repository under Settings → Git integration so the run can read the code that changed.What pays for triage and RCA
Your organization starts with a model credit that pays for triage and RCA. It is granted once and does not renew. When it is used up, add your own model provider key under Settings → Providers, and triage and RCA run on that key from then on. Nothing else depends on the credit: traces keep arriving, classifiers keep running, and findings keep being filed. A key you add before the credit runs out is used first. Usage and credit covers the credit and your limits in detail.Read next
Usage and credit
What counts as a trace, how stored data is measured, and what happens at each limit.
Data handling
What happens to your traces and credentials on Tessary Cloud, and how to delete your data.
How Tessary works
The four stages in detail, from trace to explained case.
Cloud and self-hosted
What differs between Tessary Cloud and a self-hosted open edition.