Capabilities dashboard

Use this page when you want to see every skill, tool, and connector across all your runtimes in one place, understand why a tool is greyed out, and turn the manageable ones on or off. The dashboard is the human-facing surface over the unified capability inventory, the same merged stream the Ghost Graph reads. It groups every runtime’s capabilities by runtime, then by kind, and renders a per-row action set that is a pure function of the row’s manageability tier: the UI can never offer an action the owning runtime forbids. The panel is CapabilitiesPanel, backed by GET /api/capabilities and POST /api/capabilities/:action.

​

Prerequisites

• At least one runtime to read capabilities from. clawboo-native is built in and always contributes rows; connecting openclaw, claude-code, codex, or hermes adds their capabilities. See Connecting runtimes.
• Nothing else: reads are non-destructive, and a disconnected source degrades to its last-known rows rather than blanking the inventory.

​

Where it lives

Open Settings (the gear at the bottom of the sidebar, or Cmd/Ctrl + ,), then Capabilities under the Workspace group (the puzzle-piece icon). There is no Cmd/Ctrl number shortcut for this view. The header shows a live count pill (N capabilities · M runtimes) and a Refresh button. The diagnostics drawer elsewhere in the app can also deep-link here with a runtime pre-selected.

​

How to read the inventory

​

Groups: runtime → kind

Rows are grouped by the runtime that owns the capability, in a fixed order (clawboo-native, then openclaw, claude-code, codex, hermes, human). Each group header shows the runtime label and a count. Within a group, rows sort by kind, then alphabetically by name. There are three kinds, each with its own icon: When more than one runtime contributes capabilities, a row of filter pills appears above the groups: All runtimes plus one pill per runtime. Click a pill to narrow the view to a single runtime.

​

Availability badges

Every row carries a StatusPill derived from the record’s availability and status: An unavailable row is rendered greyed (reduced opacity plus a grayscale filter) because its declarative availability requirement, an auth credential, a config value, an env var, or a plugin, is not satisfied. The server evaluates availability into available plus a diagnostics list; an unavailable row never offers an Enable/Disable button, so the action set reads consistently with the greyed presentation.

​

Manageability-derived actions

The action on the right of each row is a pure function of the row’s manageability tier, never a per-runtime literal in the UI: A few rows look manageable by tier but still render no button, by design:

• An observe-only row shows built-in, managed by <runtime> with a lock or eye-off icon and no action.
• An unavailable (greyed) row offers no action even if its tier is managed; its requirement is unmet.
• A runtime-of-record row whose owning source can’t yet write it (the source sets writable: false, e.g. an OpenClaw connector or plugin whose config.patch write is a follow-up) renders no dead button.
• A manageable-but-pending-auth row renders a disabled Enable button carrying the source’s hint; for a Codex connector the hint is pending auth — run \codex login“.

​

Steps

​

Enable or disable a capability

1. Open Capabilities from the sidebar.
2. Find the row. Use a runtime filter pill or scroll to its group.
3. Confirm the action is offered: the row has an Enable (mint) or Disable (neutral) button. If it shows built-in, managed by <runtime>, it is observe-only and cannot be changed here.
4. Click the button. The row goes busy (a spinner replaces the action icon) while POST /api/capabilities/enable (or /disable) runs with { id }.
5. On success the panel refreshes and re-reads the inventory, so the pill flips (Disabled ↔ Ready).

If the action is rejected, for example a 422 because the row is observe-only or not writable, a toast surfaces the typed error, and the inventory is not refreshed because nothing changed.

​

Refresh / handle a degraded source

• Click Refresh in the header to re-read every source.
• If a source is degraded (a disconnected Gateway, a missing runtime home), a warning banner names the source and its reason, and the dashboard shows that source’s last-known capabilities rather than dropping them. Reconnect the source, then Refresh.

​

How it’s wired

GET /api/capabilities accepts optional runtime, kind, scope, and agentId query filters and returns one merged view { records, sources }. The service fans the source adapters, persists each healthy source’s rows (source-scoped reconcile), serves last-good rows for any degraded source, and de-dupes fresh-wins by id. The browser client treats a failed fetch as ok: false so a total failure reads as an error with a Retry, not as an empty inventory. POST /api/capabilities/:action handles enable, disable, install, and approve. The server re-resolves the target row and gates enable/disable the same way the UI does: an observe-only or non-writable row returns 422; a missing row returns 404. The approve action reuses the existing tool-approval handshake; there is no second approval path.

​

Options / variations

​

Verify it worked

• After an Enable/Disable, the row’s pill flips between Ready and Disabled and the header count is unchanged.
• A direct GET /api/capabilities returns the row with the updated status.
• A rejected action raises a toast and leaves the inventory untouched; that is the expected response when you try to act on an observe-only or non-writable row.

​

Troubleshooting

​

Related

• Capabilities (concept), the inventory model and manageability tiers
• /api/capabilities reference, full request/response shapes
• Connecting runtimes, bring runtimes online so they contribute capabilities
• Marketplace, install curated skills
• Approvals, the tool-approval queue embedded in this panel

​

See also

• Ghost Graph, reads the same capability stream as per-agent skill nodes
• Runtimes
• Glossary