What is this?
Grafana Agent Observability monitors AI agents — the ones you build and the coding agents you use. This repo holds the open-source SDKs and plugins that send telemetry to it.
Grafana Agent Observability monitors AI agents — the ones you build and the coding agents you use. This repo holds the open-source SDKs and plugins that send telemetry to it.So is it a cloud dashboard, or code that lives in my project?Both. Dashboards and analytics live in your Grafana Cloud stack, while this repo is the client side: SDKs for five languages plus the agento11y CLI and plugins that ship data up to it.
Both. Dashboards and analytics live in your Grafana Cloud stack, while this repo is the client side: SDKs for five languages plus the agento11y CLI and plugins that ship data up to it.How do I start watching my coding agent?Install the agento11y CLI, run agento11y login, then launch your agent through it — for example agento11y codex. The Coding Agent Observability guide lists all install options and supported agents.
Install the agento11y CLI, run agento11y login, then launch your agent through it — for example agento11y codex. The Coding Agent Observability guide lists all install options and supported agents.What do I need before data starts flowing?A Grafana Cloud stack: take the API URL, Instance ID, API token and OTLP endpoint from the plugin's Connection tab, and create one token scoped with sigil:write, metrics:write, traces:write, and logs:write.
A Grafana Cloud stack: take the API URL, Instance ID, API token and OTLP endpoint from the plugin's Connection tab, and create one token scoped with sigil:write, metrics:write, traces:write, and logs:write.What is the most common way setup silently fails?The SDKs do not create OpenTelemetry providers themselves — your app must register them. Skip that and conversations still arrive, but analytics stays empty.
The SDKs do not create OpenTelemetry providers themselves — your app must register them. Skip that and conversations still arrive, but analytics stays empty.How much of my agent's content does it capture?The SDKs ship full generation messages but keep tool-execution arguments and results out of spans by default; the coding-agent plugins go further and capture metadata only.
The SDKs ship full generation messages but keep tool-execution arguments and results out of spans by default; the coding-agent plugins go further and capture metadata only.What will I actually see about my coding agents?Usage, cost, token spend and tool calls for every captured session, across Cursor, Claude Code, Codex and the rest, joined in one place instead of scattered per-agent logs.
Usage, cost, token spend and tool calls for every captured session, across Cursor, Claude Code, Codex and the rest, joined in one place instead of scattered per-agent logs.Can an AI coding agent wire the SDK into my app for me?Yes. Install Grafana's gcx CLI (quick script or Homebrew), then the agento11y-instrument skill: the agent inspects your code, applies small SDK diffs after your confirmation, and verifies the data lands in Grafana Cloud. Agents without skill support can be given llms.txt instead, or you can use the onboarding wizard in the Grafana Cloud plugin.
Yes. Install Grafana's gcx CLI (quick script or Homebrew), then the agento11y-instrument skill: the agent inspects your code, applies small SDK diffs after your confirmation, and verifies the data lands in Grafana Cloud. Agents without skill support can be given llms.txt instead, or you can use the onboarding wizard in the Grafana Cloud plugin.What are guards?Rules the SDK evaluates on your request path before a provider call: a guard can allow the request, deny it, or return transformed input such as redacted messages. Denied preflight attempts still show up in conversation detail.
Rules the SDK evaluates on your request path before a provider call: a guard can allow the request, deny it, or return transformed input such as redacted messages. Denied preflight attempts still show up in conversation detail.Can it grade my agent's quality automatically?Yes. Examples include offline experiments — run an agent over a dataset, grade it, publish results — and evaluators can score live traffic, so teams catch regressions without reading every conversation. There are CI quality-gate examples too.
Yes. Examples include offline experiments — run an agent over a dataset, grade it, publish results — and evaluators can score live traffic, so teams catch regressions without reading every conversation. There are CI quality-gate examples too.What are similar projects?overmind, mlflow, bifrost, signoz, casbin-gateway, phoenix, ThinkWatch-Lite, opik
overmind, mlflow, bifrost, signoz, casbin-gateway, phoenix, ThinkWatch-Lite, opikWhich topics is it in?agent-observability, cost-tracking, tracing, grafana, telemetry, opentelemetry, llmops, metrics, langchain, csharp, observability, gemini
agent-observability, cost-tracking, tracing, grafana, telemetry, opentelemetry, llmops, metrics, langchain, csharp, observability, gemini