v0.3.12: Claude Code native MCP tool calling

What's Changed

Claude Code now uses its native MCP integration instead of a custom tool_call format. The dashboard passes connected MCP server configs directly to claude --mcp-config with --permission-mode bypassPermissions, giving Claude direct tool access.

This means Claude Code can now execute MCP tools (query databases, list schemas, etc.) without asking for permission.

Full Changelog: v0.3.11...v0.3.12

Full Changelog: v0.3.11...v0.3.12

v0.3.11: All MCP servers included by default

What's Changed

All connected MCP servers and their tools are now automatically included in the chat context. No need to manually toggle each server — just start chatting and all tools are available.

You can still use Tab/Space to select specific servers if you want to limit context.

Full Changelog: v0.3.10...v0.3.11

Full Changelog: v0.3.10...v0.3.11

v0.3.10: Improve subprocess tool calling reliability

What's Changed

Strengthened tool calling instructions in the subprocess system prompt. Models were treating MCP tool definitions as informational context and refusing to call them. The prompt now explicitly instructs the model that it has direct access and must use the tools.

Full Changelog: v0.3.9...v0.3.10

Full Changelog: v0.3.9...v0.3.10

v0.3.9: MCP tool calling for Cursor and Claude Code

What's Changed

Subprocess providers (Cursor, Claude Code) can now call MCP tools through the dashboard's connected servers.

How it works

1. The system prompt includes tool definitions and instructs the AI to use ```tool_call blocks
2. The dashboard parses tool call blocks from the response
3. Tools are executed via the dashboard's already-connected MCP servers
4. The subprocess is re-invoked with results appended — same agentic loop as API providers

Example

Ask "List all schemas" with the Cursor or Claude Code provider and a db-tunnel MCP server connected — the AI will call list_schemas and show results.

Full Changelog: v0.3.8...v0.3.9

Full Changelog: v0.3.8...v0.3.9

v0.3.8: Interactive model switching for Claude Code and Cursor

What's Changed

Press m in the Chat tab to cycle through models:

Claude Code models

claude-code (default) → sonnet → opus → haiku

Cursor models

auto (default) → claude-4.6-opus-high-thinking → claude-4.6-opus-high → claude-4.6-sonnet-medium-thinking → claude-4.6-sonnet-medium → gpt-5.4-medium → gpt-5.4-high → gemini-3.1-pro

The selected model is displayed in the status bar and passed via --model flag to the subprocess.

Full Changelog: v0.3.7...v0.3.8

Full Changelog: v0.3.7...v0.3.8

v0.3.7: Model selection for Cursor and Claude Code providers

What's Changed

Added configurable model selection for subprocess providers.

Cursor

Set model in your ai.json to any cursor-agent model (run cursor-agent models to see the full list):

"cursor": {
  "command": "cursor-agent",
  "args": ["--print", "--trust"],
  "model": "claude-4.6-opus-high-thinking"
}

Default is "auto" which lets cursor-agent choose.

Claude Code

The model field is also available for Claude Code config, displayed in the status bar.

Full Changelog: v0.3.6...v0.3.7

Full Changelog: v0.3.6...v0.3.7

v0.3.6: Add --trust flag for cursor-agent

What's Changed

Added --trust to default cursor-agent args so it works in non-interactive subprocess mode without requiring manual workspace trust confirmation.

Full Changelog: v0.3.5...v0.3.6

Full Changelog: v0.3.5...v0.3.6

v0.3.5: Better subprocess error messages & Cursor API key support

What's Changed

• Better error messages: Subprocess providers (Claude Code, Cursor) now show the actual error (e.g. "Authentication required. Please run 'agent login' first") instead of a generic "exited with error"
• API key support: Added api_key field to subprocess config — set your CURSOR_API_KEY in ~/.config/mcp-dashboard/ai.json and it will be passed automatically
• CI fix: Removed unused fields that caused clippy failure

Example ai.json cursor config

"cursor": {
  "command": "cursor-agent",
  "args": ["--print"],
  "api_key": "your-cursor-api-key-here"
}

Full Changelog: v0.3.4...v0.3.5

Full Changelog: v0.3.4...v0.3.5

v0.3.4: Fix Cursor provider to use cursor-agent CLI

What's Changed

Bug Fix: Cursor provider now uses cursor-agent --print instead of cursor --chat (which opened a GUI window instead of providing CLI chat).

Changes

• Default Cursor command changed from cursor --chat to cursor-agent --print
• Both Claude Code and Cursor now use the -p flag for prompt passing
• JSON result extraction works for both providers

Full Changelog: v0.3.3...v0.3.4

Full Changelog: v0.3.3...v0.3.4

v0.3.3: Fix AI chat JSON display for subprocess providers

What's Changed

Bug Fix: Claude Code and Cursor chat providers were displaying raw JSON responses instead of the human-readable message text.

Changes

• Subprocess chat output is now buffered and parsed — the result field is extracted from JSON responses
• Token usage (input_tokens, output_tokens, cache tokens) is extracted from JSON for accurate tracking
• Non-JSON output (plain text) continues to work as-is
• Default Claude Code args simplified to --print (removed --output-format json)

Files Modified

• src/chat_provider.rs — JSON parsing in subprocess_chat
• src/chat_config.rs — Simplified default Claude Code args

Full Changelog: v0.3.2...v0.3.3

Full Changelog: v0.3.2...v0.3.3