[FR] Bulk-assign local MCP servers across agents (apply-to-all / select-all) #884

Problem

Settings → MCP → Local MCP assigns servers one agent at a time. Each server card carries a per-app checkbox row, and every tick is an individual write. The cost scales as O(servers × agents):

• adding one new agent to an existing setup means ticking every server card once
• adding one new server means ticking every agent card once

With a double-digit number of local servers and several enabled agents, that is dozens of clicks for what is conceptually a single decision ("this set of servers belongs to my whole toolbox").

Request

Any bulk affordance would help, e.g.:

• an "Apply to all agents" / "Apply to all enabled agents" action on a server card
• a "select all / clear" control on the per-app checkbox row
• the inverse direction: a per-agent "assign all local servers" action on that agent's own settings page

A confirmation step is fine; a preview of which agents would end up unassigned (see below) would be better.

Implementation context — why this is not just a button

Assignments are not stored centrally. The scan in src-tauri/src/commands/mcp.rs merges per-agent readers (local_mcp_readers → scan_local_servers_from_readers), so a server's Enabled Apps row is derived from the presence of that entry in each agent's own config file — ~/.claude.json, ~/.codex/config.toml, ~/.grok/config.toml, $DSH_HOME/mcp.json, … 15 apps in ALL_MCP_APPS, several unrelated file formats (JSON, TOML, YAML, gemini-schema, Cline's private shape).

Two consequences a bulk action has to respect:

1. mcp_set_server_apps / mcp_upsert_local_server mean "these apps and NO others" — every agent left out of the list gets the server removed. A bulk apply is therefore N upserts plus (M−N) removals, not N upserts.
2. app_can_host_spec excludes pairs an app cannot represent — e.g. SSE is rejected for Codex, and for DeepSeek the bridge explicitly rejects it (dsh-mcp-client has only stdio and streamable-http, and a rejected server fails session/new rather than being skipped). A blanket "apply to all" that ignores this would break sessions instead of degrading gracefully.

require_complete_scan already guards the analogous case on the single-agent path: if any agent's config cannot be read, the write is refused rather than acting on an incomplete picture. A bulk action must keep that guarantee — it matters more there, since one unreadable file would otherwise turn a bulk apply into a mass-removal.

Related

• [FR] Add MCP servers to custom agent via ACP #699 — custom agents have no Enabled Apps entry at all, so they are outside the scope of any bulk action until that lands.