Skip to main content
If your agent sends OpenTelemetry traces to Dash0, connect the dataset to your Mutagent workspace once, and Mutagent pulls the traces from it. Dash0 stays where your traces live. Mutagent reads them with a token that can only read.

Your agent → Dash0

Your agent exports spans and logs straight to Dash0 over OTLP. You don’t need a collector.

Dash0 → Mutagent

A Dash0 source in your workspace pulls the traces for Evaluate and Diagnose.
Helix on your machine can’t read Dash0 directly yet. Connect the source to your workspace and pull, save the result with pull-traces --out traces.jsonl, and read that file with a local JSONL source (format: unitf).

Before you start

  • A Dash0 dataset that your agent’s traces go to. A dedicated dataset keeps them apart from your other telemetry.
  • A read token for Mutagent. In Dash0, open Settings → Auth Tokens and create a token with the Reading permission. Under Dataset, pick Specific datasets and select that dataset. Don’t give Mutagent a token that can ingest or configure.
  • Your regional API URL, from Settings → Endpoints → API, for example https://api.eu-west-1.aws.dash0.com.
The harness that sends the traces needs its own token with the Ingesting permission, scoped to the same dataset. See Send traces from Claude Code. Dash0 shows a token’s value only once, when you create it.

Connect the source

With the CLI

You need the mutagent CLI, signed in, with a workspace selected (see Sign in). Keep the token off the command line: export it in your own terminal and pass the variable’s name with --secret-env.
The token is write-only: no command output ever shows it. test exits 0 when Mutagent can read the dataset, and 1 when it can’t, naming the check that failed: credentials_refused, unreachable, http_error, invalid_response or host_refused. To replace the token later, run mutagent integrations sources rotate <source-id>.

In the web app

Open Configuration › Integrations and add a trace source. Pick Dash0, then fill in the form:
  • API endpoint: your regional API URL.
  • Dataset: leave it empty for your organization’s default dataset.
  • Auth token: the read token.
Click Save and test. Mutagent stores the token encrypted and tests the connection straight away.

Pull traces

A pull reads the source’s traces into your workspace. It starts at once and runs on the server:
Poll pull-get until the pull’s status is no longer running. Then read the result with pull-manifest (what the pull holds and how complete it is) and pull-traces (the traces as JSONL, to stdout or to a file with --out). A source runs one pull at a time.

Send traces from Claude Code

Claude Code can export straight to Dash0’s OTLP endpoint. Set these variables in the environment Claude Code starts in: in your shell, or in the env block of your user settings (~/.claude/settings.json) or managed settings. A project’s .claude/settings.json can’t turn telemetry on.
  • Endpoint: your region’s OTLP ingress, from Settings → Endpoints, with no path.
  • Headers: the ingest token and the dataset to write to. The Dash0-Dataset header works for sending only: Mutagent’s reads put the dataset in the request instead, from the source’s settings.
  • gen_ai.agent.name: the name a pull’s --agent-name matches. Claude Code copies the resource attributes onto every span and log, which Dash0 needs to filter on them. Without it, a trace is matched by its service.name.
Claude Code sends no conversation text unless you turn it on. These flags add it:
Claude Code also has OTEL_LOG_TOOL_CONTENT=1, which sends tool output (file contents, command output, web results). Mutagent doesn’t read it, so leave it off.
With these flags on, prompts, file paths, commands, replies and (with OTEL_LOG_TOOL_CONTENT) tool output land in Dash0 and in every pull. Turn them on only for runs whose content may leave your machine, and send them to a dataset whose access you control.
Traces show in Dash0 shortly after a run ends. Agents that run in Mutagent Cloud don’t send to Dash0.

Filters and limits

  • Size of a pull. At most 5,000 traces; without --limit the server reads 500. With --agent-name or --tag, a pull checks at most 1,000 matching traces, whatever --limit says; narrow the window to read more. A pull that hits a limit, is throttled by Dash0 or fails partway still succeeds, marked truncated with complete: false in its manifest.
  • Large traces. Dash0 returns at most 2,000 spans and 200 logs per trace in one read and doesn’t say when it cut a trace. When a trace reaches either count, Mutagent reads it again page by page, up to 10,000 spans and 10,000 logs. Past that the trace is kept partial and the pull is marked truncated.
  • What Mutagent reads from Claude Code. Mutagent doesn’t read tool output text, thinking or compaction from a Claude Code trace: a tool call shows its input, never its result.
  • Ingest window. Dash0 accepts data timestamped within about 24 hours of now. You can’t send older runs later.
  • Retention. Dash0 keeps data for a limited time that depends on your plan. Pull what you want to keep before it ages out.

Troubleshooting

The token is limited to one dataset, but the request read the default one. Set --dataset (or the Dataset field) to the dataset the token is scoped to.
Open the dataset in the Dash0 UI and check that it holds traces in the window you pulled. Most often the harness writes to a different dataset than the source reads, or the runs fall outside the window. Without --since, a pull reads only the last 24 hours.
Dash0 answers 200 even when it stores nothing. Check that the harness’s token has the Ingesting permission (a token with only Reading can’t write) and that the Dash0-Dataset header names the dataset you’re looking at. Data timestamped more than about 24 hours away from now is dropped too.
Dash0 refused the token. If the token is scoped to one dataset, check that the source names it (--dataset or the Dataset field). Otherwise check that it has the Reading permission, then replace it with mutagent integrations sources rotate <source-id>.