AI Access (MCP)

What AI Access does

AI Access lets AI coding tools "manage your servers over a local, token-secured MCP endpoint. Credentials never leave Server Compass." Open the AI MCP tab from the sidebar. The status pill reads Running · port {n}, Enabled, not running, or Disabled, and a toggle turns access on or off.

Create a token and choose tiers

The tab has Setup, Tools, and Activity. On Setup, create a token (the default name is Claude Code), choose which permission tiers it carries, optionally restrict it to an allowlist of servers, then copy the install snippet. Permission tiers:

• Read — list servers, apps, logs, and templates; always safe and always on
• Operate — start, stop, restart, and redeploy apps
• Deploy — deploy templates, add servers, and manage domains, env vars, and secrets
• Danger — destructive actions: delete apps or secrets and run raw SSH commands

Supported clients

AI Access generates install snippets for Claude Code, Codex CLI, Grok CLI, Gemini CLI, OpenCode, Cursor, Windsurf, Cline, and VS Code. The default endpoint is http://127.0.0.1:52700/mcp, the server registers as servercompass (short alias sc), and its tools appear to the AI as mcp__servercompass__*. Token secrets are stored in your operating system keychain, and you can revoke a token at any time.

Long-running commands

run_command (Danger tier) waits up to timeoutSeconds: 120 seconds by default, 900 at most. For restarts, builds or anything with a sleep in it, the agent passes background: true. The call returns a job id straight away and the command keeps running on the server. The agent then calls get_command_job (Read tier) with that id to see whether the job is still running, its exit code, and the last lines of output (200 by default, up to 1,000). Job files live in ~/server-compass/mcp-jobs on the server and are cleaned up after seven days.

When a call is denied

If a token lacks the tier a tool needs, the agent is told which tier is missing and to create a new token in AI MCP → Access tokens, since tokens cannot be edited. The message also says that this is a Server Compass token setting, not one of the agent's own permission settings. A command that matches the destructive-command denylist is refused with the name of the rule it matched; Server Compass never runs it over MCP.

On the Activity tab, a denied call has a denied badge and a Create a token with {tier} button. It opens the token form prefilled with that token's settings plus the missing tier.

Kubernetes tools

On servers with a cluster, agents get ten Kubernetes tools on the usual tiers. Read: k8s_overview, k8s_list, k8s_logs, k8s_rollout_status, k8s_rollout_history. Operate: k8s_restart_workload, k8s_scale_workload, k8s_delete_pod, k8s_rollout_undo. Deploy: k8s_apply_manifest. Installing or uninstalling k3s is not available over MCP. See Kubernetes and k3s.

Pre-approving read tools in Claude Code

The Claude Code setup includes an optional snippet, Optional: pre-approve ServerCompass tools in Claude Code, for ~/.claude/settings.json or a project's .claude/settings.json. It lists only the read-only Server Compass tools, so Claude Code stops asking before each read. Tools that change a server still ask.