Opt-in OpenTelemetry (OTLP) export for the Agent Panel

[xmatthias]

What are you proposing?

Let users send OpenTelemetry traces, metrics and events from Zed's agent to their own OTLP collector. This must be off by default and will be completely separate from Zed's own product telemetry.

Why does this matter?

Right now there's no way to see what Zed's agent is doing, or what it costs, in the observability stack a team already runs (Grafana, Honeycomb, Datadog, Langfuse…) (intentionally not pointing to 1 particular - as that's not relevant to this, it's generally applicable)

With OTel export, it becomes easy to:

• track token usage and cost per user, team or project, especially with your own API keys
• debug agent turns: which tools ran, how long each LLM call took, and where a turn failed or retried
• meet audit and governance requirements that some organizations have before approving an AI tool
• watch your own sessions by pointing it at a local collector or Jaeger

This is even more important when not using zed's builtin AI, but relying on some third-party as AI provider (e.g. github copilot) - but zed's AI provider should also be part of this (it shouldn't discriminate).
Most other agentic tools already support this.

Are there any examples or context?

• Claude Code: CLAUDE_CODE_ENABLE_TELEMETRY=1 plus the standard OTEL_* env vars; content is redacted by default. Docs
• VS Code (Copilot Chat): github.copilot.chat.otel.* settings; follows the OTel GenAI semantic conventions. Docs
• Codex CLI: an [otel] section in config.toml; log_user_prompt = false by default. Docs
• Gemini CLI: a telemetry block in settings.json. Docs

What they have in common: off by default, standard OTLP over gRPC or HTTP, and prompt/response content redacted unless the user opts in.

Possible approach

"agent": {
  "opentelemetry": {
    "enabled": false,
    "endpoint": "http://localhost:4318",
    "protocol": "http/protobuf", // or "grpc"
    "headers": {},
    "capture_content": false     // prompts, responses, tool args/results
  }
}

• Respect the standard OTEL_EXPORTER_OTLP_* and OTEL_RESOURCE_ATTRIBUTES env vars.
• Follow the OTel GenAI semantic conventions: an invoke_agent -> chat / execute_tool span per turn, plus the gen_ai.client.token.usage and gen_ai.client.operation.duration metrics.
• Content is redacted unless capture_content is on.

Ai was used for research of this content - but every word has been read and validated by a human