Troubleshooting

Each of these looks like a broken feature and is not. They are collected here because the error each one produces points somewhere unhelpful.

Grafana returns 401 with a token that looks correct

You almost certainly pasted an access-policy token, which starts glc_. Those authorise telemetry ingestion only and return 401 against the dashboard API, which is the API used to list and render dashboards and the one the agent's tools call.

You need a service-account token, which starts glsa_. Create it in Grafana under Administration, then Service accounts, then Add token. If you signed in with OIDC you still need one: browser sign-in covers the dashboard panel, not the agent's MCP tools.

The Design engine picker appears to do nothing

A config.toml in your workspace root that names a provider takes precedence over the in-app picker, because the design engine reads the workspace file before the one in your home directory. Remove or edit that file. Changing the provider also respawns the design sidecar, so a short pause after switching is expected rather than a hang.

An MCP server will not connect

If it is a remote server reached over HTTP or SSE, it cannot be connected. MCP support here is stdio only, and remote servers are discoverable in the registry search without being connectable. Put a local stdio bridge in front of it, or use a server that ships as a local process.

The engine picker says a CLI is missing

The binary was not found on PATH. The picker shows the install command for that engine; if it is installed but not signed in, it shows the login command instead. Note that the binary name is not always the product name: Cursor CLI is cursor-agent.

The audit reports nothing useful

Scanners are not bundled, so a machine with none installed produces a manifest inventory, the agent configuration findings, and a list of tools to install. Installing osv-scanner is the single highest-value addition, since it covers every ecosystem from lockfiles.

Go modules are a special case: they are scanned by govulncheck when it is installed, andgovulncheck reports only vulnerabilities your code can actually reach, so a quiet result there is meaningful rather than empty.

Findings show an ungraded severity

Several scanners report no severity at all, including cargo audit, pip-audit and govulncheck. Those findings are counted separately in the summary rather than dropped. Ungraded means nobody classified it, not that it is minor.

Connecting Grafana or adding a connector is unavailable

Both require the bundled Lucid engine. Once a Grafana instance is connected or a connector is added, it works on every engine; only the setup step is restricted.

Design will not run on my engine

Design mode accepts Claude Code and the Codex CLI only. Cursor CLI and OpenCode drive coding tabs but have no provider behind the design canvas, so this is not a configuration you can reach.