Trace Coding Agents

Stream OpenTelemetry traces from Claude Code, Gemini CLI, Codex, GitHub Copilot, Cursor, and other coding agents into your Everstack instance — no SDK required.

Overview

Everstack ingests OpenTelemetry directly, so any coding agent that can export OTLP streams its sessions straight into your Traces view — LLM requests, tool calls, sub-agents, token usage, and cost — with no Everstack-specific SDK or instrumentation. Everstack normalizes each agent's attribute conventions into one schema, so they all render the same way.

Pick your agent

Claude Code

Anthropic's CLI. Native OTel; full trace tree in beta.

Gemini CLI

Google's CLI. Uses its own telemetry settings; traces off by default.

Codex

OpenAI's CLI. OTel configured via config.toml.

GLM

GLM (Zhipu) via Claude Code.

Kimi

Kimi (Moonshot) via Claude Code.

Cursor

Routed through the Everstack gateway (no native OTel).

GitHub Copilot

Native OTel from Copilot Chat and the Copilot SDK/CLI.

How it works

Your agent exports OTLP/HTTP to your instance. Three signals map to three endpoints, all authenticated with an Everstack API key as a Bearer token:

Signal	Endpoint
Traces	`POST https://<your-instance>/v1/traces`
Logs	`POST https://<your-instance>/v1/logs`
Metrics	`POST https://<your-instance>/v1/metrics`

The Langfuse-compatible aliases /api/public/otel/v1/{traces,logs,metrics} are also accepted. The API key is resolved to your tenant and stamped onto every span — clients cannot write into another tenant's data.

Create an API key

Open your dashboard → Settings → Vault → API Keys, create a key, and copy it (shown once). The key must be created on the same instance you send telemetry to.

Point your agent at the instance

Most OTLP exporters take a base endpoint and append the signal path automatically:

OTEL_EXPORTER_OTLP_ENDPOINT=https://<your-instance>
OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer <your-api-key>"

Then follow the page for your agent above.

What you get

Once telemetry is flowing, every agent session renders in Traces with a unified schema, regardless of which agent produced it:

Model & provider — normalized across agents; provider is inferred from the model name when it isn't reported (e.g. glm-* → Zhipu, kimi-* → Moonshot).
Token usage — input, output, and cached tokens, coalesced across each agent's attribute names.
Cost — reported where available; otherwise computed from your model-catalog pricing.
Tool calls & sub-agents — tool spans and nested sub-agent runs, rendered in the trace tree.

Cost for token-only agents. Only some agents (e.g. Claude Code) report a dollar cost. For agents that emit only token counts (Gemini CLI, Codex, GLM, Kimi), Everstack computes cost from the same model-catalog pricing used for billing — so trace cost matches what you are billed. A reported cost is never overwritten.

Verify

Send a prompt through your agent, then open Traces. Traces appear within a few seconds (metrics aggregate more slowly). Filter by service name or session.id to find a specific run; the root span expands into LLM requests, tool calls, and sub-agents.

Telemetry initializes when the agent process starts, so only sessions started after you add the config will emit. Restart the agent — a new prompt in an existing session is not enough.

Notes

The API key is sent as a Bearer token and must belong to the instance you point at — a key from another instance will not authenticate.
To expose your deployed agents as callable tools to an IDE or another agent, see the MCP server — that's the inbound direction (others calling your instance), complementary to tracing.