Grafana Labs has released two tools that allow AI coding agents to access real-time observability data while writing code: the gcx command-line interface and the Grafana MCP server. Both products pull metrics, logs, traces, service-level objectives, and synthetic monitoring results from either Grafana Cloud or self-hosted installations. The company announced the tools at GrafanaCon 2026 in April and has now made them generally available, aiming to connect the pace at which agents produce code with engineers' grasp of what that code actually does in production.
The two offerings target different levels of structure. The Grafana MCP server delivers a predetermined collection of tools designed for typical tasks and runs either on a user's own infrastructure or via a hosted endpoint for Grafana Cloud deployments. The gcx CLI provides a more adaptable option that agents can use to construct custom workflows, compatible with Grafana Cloud or self-hosted open-source and enterprise versions. Grafana Labs also distributes an installable skill bundle for gcx and separate Claude Code plugins for both the MCP server and a Grafana Assistant guidance layer. In a practical workflow, the agent can ground its implementation choices in actual system behavior instead of assumptions from training data. When adding a new payment provider, for instance, the agent can check current RED metrics and discover that the p95 latency for the existing provider sits at two seconds, then use that figure to configure mocked latency in unit and integration tests and estimate request rates for the new handler. Agents can read and modify existing dashboards, trace dashboard queries back to the code generating the telemetry, and push revised dashboard definitions into Grafana or version control.
According to Grafana Labs, the tools address a frequent issue in agentic coding pipelines: developers can plan, build, and review changes largely through a large language model, yet they may still approve a pull request without fully understanding it. Historically, engineers developed a mental model of a change as they typed each line. When an agent writes the diff instead, that model-building stage vanishes, even as review agents and automated approvals generate a misleading sense of assurance. The company frames its solution as bringing observability data into the workflow to serve as an extra, evidence-driven checkpoint on agent output rather than depending only on code review. Dafydd-t, a Grafana Labs employee, explained on Reddit that "the Grafana MCP server provides more opinionated tools for common tasks when interacting with Grafana" while "gcx provides a broader, less opinionated set of tools that you can use to build your own workflows."
For local iteration, Grafana Labs notes that agents already understand OpenTelemetry tooling: they can spin up an OpenTelemetry Collector and send local build telemetry to a Grafana stack, or use the grafana/otel-lgtm Docker image for a fully local LGTM stack, with production dashboard definitions pulled into that local instance via gcx. Traffic generation follows a similar pattern—agents can use production telemetry to write realistic k6 load-test scripts and set up k6 in a container, work that previously required roughly a day of manual effort. The k6 tool includes its own agent-skill bundle, installable through k6 x agent init, which can author tests from observed traffic and help distinguish a failing test from a failing system under test. Grafana Labs shares an internal case: the Tempo squad uses a custom agentic harness to profile a development environment holding several terabytes of data, identify optimization candidates, run the same baseline query after implementing changes to measure the difference, and then let a person decide whether to merge. The company has also launched an experimental Agentic Testing feature that uses natural-language instructions to validate UI flows in a live web application, addressing the gap left by backend telemetry that doesn't catch front-end regressions. A pull request now looks different: instead of only an agent-generated diff and description, it can include dashboard links showing the local build under realistic load, with new metric or log labels visible right when a test ran. For engineering teams running agentic coding pipelines, the tools primarily introduce a verification layer that ties agent output to actual system behavior, shifting the question from "did the agent understand the ticket?" to "does the running system match our expectations?"—a distinction that matters as review agents increasingly approve changes without close scrutiny. Organizations that lean heavily on agent-authored code may find the boundary between automated assistance and genuine oversight harder to defend without instrumentation anchored in observable system state.