Skip to content

P0: Shared custom MCP server manager via UI (HTTP + stdio, all agents) #5

Open

Description

@jameswright1562

Goal

Settings → Integrations → MCP Servers: add custom servers through UI, once, and select per-agent access (Claude Code, Codex, Cursor and OpenCode/custom models).

Requirements

  • Create/edit/disable/delete/test servers; HTTP Streamable and local stdio command/args/cwd.
  • OAuth login/refresh for remote providers, static auth headers, tokens saved in macOS Keychain/secret store, never in prompts/logs/agent env unless explicitly necessary.
  • Per-project/per-agent enablement, tool allowlist, on-demand tool discovery and resilient reconnect.
  • A nonworking MCP server must not crash another native agent's session.
  • T3 server owns shared connections; an external agent receives only appropriately scoped tool access.
  • Include connection health and tool list in UI; no manual JSON edits.
  • Test thread isolation, OAuth state, rotating credentials, reconnect, tools routing, stdio shutdown, and provider compatibility.
  • Keep distinct from T3's built-in orchestration MCP.

Reference upstream closed PR #16466 as architecture only, not an approved patch: review raised correctness and identity/credential concerns.

See draft PR #3 and architecture spec.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions