Skip to main content
eve is Vercel’s framework for durable AI agents. It has OpenTelemetry built in: every agent turn, model call, tool call and approval becomes a span that follows the GenAI semantic conventions. Bronto accepts these spans over OTLP directly, so you don’t need a Bronto SDK or a Collector. This page covers two things:

Prerequisites

  • An eve project (eve 1.0 or later)
  • A Bronto API key with the Ingestion role (see API Keys)
  • Your Bronto region: eu or us

Send traces to Bronto

Install @vercel/otel:
Create agent/instrumentation/bronto.ts:
agent/instrumentation/bronto.ts
An eve add instrumentation/bronto command that sets this up for you is on its way to eve’s integration registry.
Set two environment variables, in .env.local for eve dev and in the Vercel project’s environment variables for deployments. eve link and vercel env pull rewrite .env.local, so check the variables are still there after running either:
eve picks up every file in agent/instrumentation/, so the Bronto exporter runs next to any other destinations you have configured.

Where the traces land

Bronto stores the spans in the .traces collection. eve sets service.name to the agent’s name, so each agent’s traces are easy to tell apart without any extra configuration. To add resource attributes to every span, such as the environment, declare them in agent/instrumentation/otel.ts:
agent/instrumentation/otel.ts
eve allows one otel() declaration per project. If agent/instrumentation/otel.ts already exists, add the attributes to its resource.

What you get

Each agent turn starts a new trace:
The same export also carries the Vercel Workflow spans that run the agent’s durable session (workflow.*, step.* and queue.publish), with the same service.name. To find the agent’s own spans, filter on gen_ai.operation.name. The most useful attributes: See eve’s trace topology for the full list, and LLM Observability for searching and aggregating gen_ai.* fields in Bronto.

Control what’s recorded

The exportPolicy in bronto.ts removes prompts, replies, tool arguments and tool results before the spans leave your agent. Span names, timings, models, token counts and errors are kept. To send the content as well, for example to debug the agent’s answers, remove exportPolicy from bronto.ts. eve’s own trace policy still applies: by default it records content only in development and in conversations a channel marks as public. Change that with tracePolicy in otel().
Recorded content is searchable by everyone with access to the dataset. If conversations contain personal or regulated data, keep the redaction.

With a Vercel trace drain

eve’s exporter sends spans directly, even when the project also has a trace drain to Bronto. With both, every eve span arrives twice. Use one path for the agent’s spans. See OpenTelemetry in Vercel Functions.

Let your agent query Bronto

The Bronto MCP server gives an agent tools to search logs and traces, list datasets and run aggregations. An incident or support agent can use it to look up what happened before it answers. Create a Bronto API key with the Agent Read Only role (see API Keys). Don’t reuse the ingestion key from bronto.ts. Then create agent/connections/bronto.ts:
agent/connections/bronto.ts
Set the key in .env.local and in the Vercel project’s environment variables:
The agent finds the Bronto tools through eve’s tool search, as bronto__search_logs, bronto__get_datasets and so on. Every user of the agent sees the data the key can read.
Coming soon: sign in per user. eve can also connect to the Bronto MCP server with each user’s own Bronto login, through Vercel Connect OAuth, so the agent only sees what that user can see. This doesn’t work yet with Bronto’s MCP server. Use the API key until this page shows the OAuth setup.