Nexura
A stack trace with no map
The exception names the line that raised. Not which step was running, what it was handed, or which step before it was already wrong.
New
Trace any Python agent with the Nexura SDK
Ship AI agents you can debug in production.
Nexura records every step your agent takes — each model call, tool and retrieval, with its latency, tokens and cost — and points to the one that broke. Build on a canvas or bring your own framework, then deploy with one command.
Free to useNo credit cardBring your own model key
Works with the frameworks and models you already use
LangChain
LangGraph
OpenTelemetry
OpenAI
Anthropic
Gemini
Ollama
vLLM
The problem
When an agent fails, the hard part is finding where.
Multi-step agents fail quietly, in production, and differently every time. Logs were not built for that.
Print debugging in production
Seeing one intermediate value means adding a log line, redeploying and waiting for a failure that may never repeat the same way.
The invoice is the first alert
A retry loop burning 40k tokens a call looks like a healthy run — until the provider bill arrives at the end of the month.
Product
Everything between the prompt and the answer, on one screen.
One place to build agents, ship them, and see exactly what every run did.
Every run is a span tree
Each model call, retrieval and tool call is recorded with its parent, latency, input and output. Only the step that actually failed is marked, so it never hides inside a red column.
Cost per step, not per month
Every span is priced at the model's published rate, so an expensive step shows on the row that spent it.
Bring your own agent
Two lines trace an agent running anywhere. LangChain and LangGraph are captured automatically.
import nexura
nexura.init(service="support-agent")
agent.invoke(question) # traced
One agent, four endpoints
Every agent is callable the way your stack already works, with one key and one trace.
Change it without guessing
Every save is a version. Score it against a test set before it replaces the one in production.
For developers
From zero to your first trace in three steps.
No rewrite, no new framework. Nexura attaches to the agent you already have.
-
Create a key
Sign in, open Settings → API keys, and create a key with the “Write traces” permission.
-
Add two lines
Call
nexura.init()before your agent runs. Framework calls are captured automatically; anything else takes onewithblock. -
Open the run
Every run appears in Traces as a waterfall with latency, tokens and cost — and the failed step selected for you.
The Python SDK is in preview. Other languages can send traces to the same HTTP endpoint.
Managed hosting
Deploy with one command. Trace from the first request.
Push a Python service and Nexura builds it, runs it, and connects every revision to its traces. No Dockerfile, no cloud console.
- Isolated per workspace. Each workspace gets its own cloud project, build identity and runtime identity.
- Secrets stay secret. Provider keys go to a secret manager and never into your image.
- Roll back in one click. Every deploy is an immutable revision you can route traffic back to.
- Private by default. Calls go through an authenticated proxy, never an open URL.
Terminal
$ nexura init Wrote nexura.json $ nexura secret support-agent OPENAI_API_KEY Value for OPENAI_API_KEY: •••••••• $ nexura deploy --yes Source bundle: 4 files, 12,480 bytes queued building deploying live https://nx-support-agent.a.run.app
Managed hosting is in preview and enabled per workspace. Tracing works everywhere today.
134Providers reachableBy model string, no allowlist
116Models in the pickerValidated, with tool-calling flags
1Credential surfaceNot one SDK per vendor
0%Markup on inferenceBilled by your provider, to your key
FAQ
Questions people ask before signing up.
Do I have to build my agent in Nexura to trace it?
No. Agents built in the studio are traced automatically. Agents running anywhere else report through the
Python SDK: LangChain and LangGraph are captured with no code changes, OpenTelemetry spans are collected,
and any other Python code needs one with nexura.trace() block. Other languages
can post to the same HTTP endpoint.
Can Nexura host my agent for me?
Yes, in preview. nexura deploy uploads a Python HTTP service, builds it and runs
it in a cloud project isolated to your workspace. Each revision is traced and can be rolled back. Each
account can run up to 5 services and start 20 builds a day. Hosting is enabled per workspace — ask us
to switch it on.
Where do my provider keys live?
In an encrypted secret store scoped to your workspace. They are used to call the provider and are never returned by the API once saved — the interface shows a prefix so you can tell two keys apart, and nothing more.
Is Nexura free?
Yes. Nexura is free to use. You pay your model provider directly for the calls your agents make, with your own key.
What does a run actually cost?
Whatever your model provider charges. Nexura prices each step from the model's published rate so you can see it per span, but the call is billed to your key by your provider.
Can I get my agents out?
Every agent exports as plain JSON — the same format the canvas loads. There is no proprietary container and nothing only Nexura can open.
Is this built on Langflow?
Yes. Nexura builds on Langflow's open-source graph engine and component model, and adds the trace engine, the SDK, managed hosting, the unified model layer and this studio. Flows you already have keep loading.
Stop reading logs to find the step that broke.
Trace your first run in about ten minutes, with your own model key. Free to use.