[Feature] Support MCP Passthrough and ACP-Scoped MCP Management in ACP Mode

[zerob13]

What do you need? / 你需要什么？

Current Behavior / Problem

In ACP mode, ACP agents maintain their own MCP configuration and do not read or inherit MCP settings from DeepChat.

As a result:

• MCPs that are already configured and persisted in DeepChat (non-inmemory MCPs) are not available to ACP agents
• ACP agents can only use MCPs that are explicitly configured within their own scope
• There is currently no mechanism for DeepChat to pass MCP definitions down to ACP agents

This makes it impossible for ACP agents to reuse existing DeepChat MCP configurations without duplication, and leads to fragmented MCP management.

Feature Proposal

1. MCP Passthrough from DeepChat to ACP Agents

   • In ACP mode, allow DeepChat to pass through non-inmemory MCPs that are already recorded in DeepChat to ACP agents.
   • ACP agents should be able to load and use these MCPs without requiring separate reconfiguration.

2. Per-ACP Agent MCP Configuration

   • Provide settings to configure, for each ACP agent:

     • Which MCPs (passed from DeepChat) should be loaded

   • MCP configuration in ACP mode should be agent-scoped, not global.

3. ACP-Aware MCP State in Chat Input

   • In ACP mode, the MCP selector and enable/disable state shown in the chat input should:

     • Reflect the MCP configuration of the currently active ACP agent
     • Not reflect the global MCP state of the entire application

   • The UI should stay in sync with the ACP agent’s MCP configuration.

Expected Benefits

• Enables ACP agents to reuse existing DeepChat MCP configurations
• Avoids MCP duplication and configuration drift
• Improves consistency between agent settings and chat UI
• Makes ACP mode more intuitive and scalable for multi-agent workflows

[coderabbitai]

📝 CodeRabbit Plan Mode

Generate an implementation plan and prompts that you can use with your favorite coding agent.

✅ Successfully generated implementation plan.

💡 Use @coderabbitai <feedback> to replan

Examples

• Example 1
• Example 2

---

🔗 Similar Issues

Related Issues

• [Feature] Add ACP Support #1100
• [Feature] init process in NewThreadView #1168
• [Feature] 为 DeepChat 添加 MCP 协议的 Sampling 支持 #568
• [Feature] Support Agent Skills Standard Integration #1201

🔗 Related PRs

#554 - refactor: 将 Model Control Protocol (MCP) 重命名为 Model Context Protocol [merged]
#635 - Feature/mcp tool improve [merged]
#698 - fix: a potential problem with mcp tools [merged]

👤 Suggested Assignees

• zerob13
• A0nameless0man
• wushanchao
• sqsyli
• hllshiro

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord or schedule a call!

[coderabbitai]

Coding Plan

Approach

Enable ACP agents to use non-inmemory MCPs configured in DeepChat through a passthrough mechanism, with per-agent MCP configuration and agent-aware UI that reflects the active agent's MCP state rather than global state.

Observations

DeepChat uses a presenter-based architecture separating main process logic from renderer UI. MCPs are managed through McpPresenter with configuration stored in mcp-settings.json. ACP agents are configured via acpConfHelper with settings stored in ElectronStore at acp_agents. The conversation system supports per-conversation settings stored in chatConfig. The UI uses Pinia stores with composables for reactive state management. The codebase already has patterns for per-agent configuration (profiles, enabled state) and capability-based initialization for ACP sessions.

Assumptions

Assumption 1: Default MCP Access Behavior

Options Considered:

• Default to empty array or undefined to preserve current behavior where agents have no MCP access

Chosen Option: Default to empty array or undefined to preserve current behavior where agents have no MCP access

Assumption 2: MCP Filtering Approach

Options Considered:

• Return empty list when agent has no MCP selections
• Return all available MCPs when agent has no MCP selections

Chosen Option: Return empty list when agent has no MCP selections

Rationale: preserving current behavior

💡 User Tips

Regenerate the plan with different choices with @coderabbitai <feedback>.

Implementation Steps

Phase 1: Per-Agent MCP Configuration Storage

This phase establishes the data model for storing which MCPs each ACP agent can access, without changing runtime behavior.

Task 1: Extend Agent Configuration Schema

Add per-agent MCP selection storage to the ACP agent configuration structure.

• Extend AcpBuiltinAgent and AcpCustomAgent types in src/shared/types/presenters/legacy.presenters.d.ts to include an optional mcpSelections field (array of MCP server IDs)
• Default to empty array or undefined to preserve current behavior where agents have no MCP access

Task 2: Add Configuration Helper Methods

Create methods in acpConfHelper for managing per-agent MCP selections.

• Add methods in src/main/presenter/configPresenter/acpConfHelper.ts: getAgentMcpSelections(agentId), setAgentMcpSelections(agentId, mcpIds), addMcpToAgent(agentId, mcpId), removeMcpFromAgent(agentId, mcpId)
• Follow existing mutation helper patterns (mutateBuiltins, mutateCustoms) for safe updates
• Validate MCP IDs against available non-inmemory MCPs from McpConfHelper

🤖 Prompt for AI agents

Establish data model for storing which MCPs each ACP agent can access.

Extend AcpBuiltinAgent and AcpCustomAgent types:
- Add optional mcpSelections field (array of MCP server IDs) to type definitions
in src/shared/types/presenters/legacy.presenters.d.ts
- Default to empty array or undefined

Add configuration helper methods in
src/main/presenter/configPresenter/acpConfHelper.ts:
- Create getAgentMcpSelections(agentId) to retrieve agent's MCP selections
- Create setAgentMcpSelections(agentId, mcpIds) to update agent's MCP selections
- Create addMcpToAgent(agentId, mcpId) to add single MCP to agent
- Create removeMcpFromAgent(agentId, mcpId) to remove single MCP from agent
- Use existing mutation helper patterns (mutateBuiltins, mutateCustoms)
- Validate MCP IDs against available non-inmemory MCPs from McpConfHelper

Phase 2: MCP Tool Filtering for ACP Sessions

This phase connects agent MCP configuration to the runtime, making selected MCPs available to ACP agents through filtered tool exposure.

Task 1: Create Agent-Aware Tool Filtering Layer

Build a mechanism to filter MCP tools based on agent configuration.

• Create new helper in src/main/presenter/mcpPresenter/ (e.g., agentMcpFilter.ts) that queries agent MCP selections and filters tool lists
• Accept agent ID as input and return filtered list of MCPToolDefinitions from only the selected MCPs
• Handle case where agent has no MCP selections (return empty list, preserving current behavior)

Task 2: Integrate Filtering into ACP Session Initialization

Apply MCP filtering when ACP sessions start.

• Modify src/main/presenter/llmProviderPresenter/agent/acpSessionManager.ts or acpProvider.ts to include agent ID context during session creation
• During tool discovery phase, call the filtering layer to get agent-specific tools instead of all available tools
• Pass filtered tools through the existing ACP session initialization flow

Task 3: Enable Tool Call Routing

Ensure tool calls from ACP agents reach the correct MCPs.

• Update tool execution path in src/main/presenter/mcpPresenter/toolManager.ts to accept agent context
• Verify permission checking (checkToolPermission) respects both MCP-level and agent-level access
• Route tool calls through existing MCP client connections, maintaining the permission approval flow

🤖 Prompt for AI agents

Connect agent MCP configuration to runtime by filtering tools based on agent
selections.

Create agent-aware tool filtering:
- Create new helper file in src/main/presenter/mcpPresenter/ (e.g.,
agentMcpFilter.ts)
- Implement filtering function that accepts agent ID and returns filtered
MCPToolDefinitions
- Query agent MCP selections and filter tools to only selected MCPs
- Return empty list when agent has no MCP selections

Integrate filtering into ACP session initialization:
- Modify acpSessionManager.ts or acpProvider.ts in
src/main/presenter/llmProviderPresenter/agent/
- Add agent ID context during session creation
- Call filtering layer during tool discovery phase to get agent-specific tools
- Pass filtered tools through existing ACP session initialization flow

Enable tool call routing:
- Update tool execution path in src/main/presenter/mcpPresenter/toolManager.ts
to accept agent context
- Update checkToolPermission to respect both MCP-level and agent-level access
- Route tool calls through existing MCP client connections with permission
approval flow

Phase 3: Agent-Aware UI State Management

This phase makes the UI display MCP state for the currently active ACP agent rather than global state.

Task 1: Track Active Agent's MCP Configuration

Extend conversation state to include active agent's MCP selections.

• Add computed properties or getters in src/renderer/src/stores/chat.ts that derive active agent's MCP selections from agent configuration
• When agent changes in conversation, load that agent's MCP selections into reactive state
• Handle both ACP mode (agent-specific) and normal mode (global) contexts

Task 2: Create Agent-Scoped MCP Composables

Build UI composables that return agent-filtered MCP data.

• Extend or create new composables in src/renderer/src/components/chat-input/composables/ that check current mode (ACP vs normal)
• In ACP mode: filter mcpStore.tools, mcpStore.resources, mcpStore.prompts by active agent's selections
• In normal mode: return global MCP state (current behavior)

Task 3: Update Chat Input MCP Display

Make the chat input's MCP UI conditional on mode and active agent.

• Modify src/renderer/src/components/chat-input/ChatInput.vue to use agent-scoped composables when in ACP mode
• Update useMentionData.ts to filter suggestions based on active agent's MCPs
• Ensure MCP-related UI components (tool lists, resource viewers) respect agent context

🤖 Prompt for AI agents

Make UI display MCP state for currently active ACP agent instead of global
state.

Track active agent's MCP configuration:
- Add computed properties or getters in src/renderer/src/stores/chat.ts
- Derive active agent's MCP selections from agent configuration
- Load agent's MCP selections into reactive state when agent changes
- Handle both ACP mode (agent-specific) and normal mode (global) contexts

Create agent-scoped MCP composables:
- Extend or create composables in
src/renderer/src/components/chat-input/composables/
- Check current mode (ACP vs normal)
- In ACP mode: filter mcpStore.tools, mcpStore.resources, mcpStore.prompts by
active agent's selections
- In normal mode: return global MCP state

Update chat input MCP display:
- Modify ChatInput.vue in src/renderer/src/components/chat-input/ to use
agent-scoped composables in ACP mode
- Update useMentionData.ts to filter suggestions based on active agent's MCPs
- Ensure MCP-related UI components (tool lists, resource viewers) respect agent
context

Phase 4: Agent MCP Configuration UI

This phase adds the user interface for selecting which MCPs each agent can access.

Task 1: Create MCP Selection Component

Build a reusable component for agent MCP configuration.

• Create new component in src/renderer/src/components/mcp-config/ (e.g., AgentMcpSelector.vue)
• Display available non-inmemory MCPs as checkboxes or toggles
• Accept agent ID as prop and load that agent's current selections
• Emit selection changes to parent component

Task 2: Integrate into Agent Settings

Add MCP configuration to agent settings panels.

• Locate agent configuration UI (likely in src/renderer/src/components/ or settings directory)
• Embed the MCP selection component in each agent's settings view
• Wire up save actions to call IPC methods that invoke acpConfHelper.setAgentMcpSelections()

Task 3: Add Agent MCP Status Indicators

Provide visual feedback about agent MCP configuration.

• Add indicators in agent selection UI showing how many MCPs each agent has configured
• Update agent cards/rows to display MCP count or status badge
• Consider adding tooltips listing configured MCP names

🤖 Prompt for AI agents

Add user interface for selecting which MCPs each agent can access.

Create MCP selection component:
- Create new component in src/renderer/src/components/mcp-config/ (e.g.,
AgentMcpSelector.vue)
- Display available non-inmemory MCPs as checkboxes or toggles
- Accept agent ID as prop and load agent's current selections
- Emit selection changes to parent component

Integrate into agent settings:
- Locate agent configuration UI in src/renderer/src/components/ or settings
directory
- Embed MCP selection component in each agent's settings view
- Wire up save actions to call IPC methods that invoke
acpConfHelper.setAgentMcpSelections()

Add agent MCP status indicators:
- Add indicators in agent selection UI showing MCP count per agent
- Update agent cards/rows to display MCP count or status badge
- Add tooltips listing configured MCP names

---

🚀 Next Steps

🤖 All AI agent prompts combined

Task: 1

Establish data model for storing which MCPs each ACP agent can access.

Extend AcpBuiltinAgent and AcpCustomAgent types:
- Add optional mcpSelections field (array of MCP server IDs) to type definitions
in src/shared/types/presenters/legacy.presenters.d.ts
- Default to empty array or undefined

Add configuration helper methods in
src/main/presenter/configPresenter/acpConfHelper.ts:
- Create getAgentMcpSelections(agentId) to retrieve agent's MCP selections
- Create setAgentMcpSelections(agentId, mcpIds) to update agent's MCP selections
- Create addMcpToAgent(agentId, mcpId) to add single MCP to agent
- Create removeMcpFromAgent(agentId, mcpId) to remove single MCP from agent
- Use existing mutation helper patterns (mutateBuiltins, mutateCustoms)
- Validate MCP IDs against available non-inmemory MCPs from McpConfHelper
===============================================================================

Task: 2

Connect agent MCP configuration to runtime by filtering tools based on agent
selections.

Create agent-aware tool filtering:
- Create new helper file in src/main/presenter/mcpPresenter/ (e.g.,
agentMcpFilter.ts)
- Implement filtering function that accepts agent ID and returns filtered
MCPToolDefinitions
- Query agent MCP selections and filter tools to only selected MCPs
- Return empty list when agent has no MCP selections

Integrate filtering into ACP session initialization:
- Modify acpSessionManager.ts or acpProvider.ts in
src/main/presenter/llmProviderPresenter/agent/
- Add agent ID context during session creation
- Call filtering layer during tool discovery phase to get agent-specific tools
- Pass filtered tools through existing ACP session initialization flow

Enable tool call routing:
- Update tool execution path in src/main/presenter/mcpPresenter/toolManager.ts
to accept agent context
- Update checkToolPermission to respect both MCP-level and agent-level access
- Route tool calls through existing MCP client connections with permission
approval flow
===============================================================================

Task: 3

Make UI display MCP state for currently active ACP agent instead of global
state.

Track active agent's MCP configuration:
- Add computed properties or getters in src/renderer/src/stores/chat.ts
- Derive active agent's MCP selections from agent configuration
- Load agent's MCP selections into reactive state when agent changes
- Handle both ACP mode (agent-specific) and normal mode (global) contexts

Create agent-scoped MCP composables:
- Extend or create composables in
src/renderer/src/components/chat-input/composables/
- Check current mode (ACP vs normal)
- In ACP mode: filter mcpStore.tools, mcpStore.resources, mcpStore.prompts by
active agent's selections
- In normal mode: return global MCP state

Update chat input MCP display:
- Modify ChatInput.vue in src/renderer/src/components/chat-input/ to use
agent-scoped composables in ACP mode
- Update useMentionData.ts to filter suggestions based on active agent's MCPs
- Ensure MCP-related UI components (tool lists, resource viewers) respect agent
context
===============================================================================

Task: 4

Add user interface for selecting which MCPs each agent can access.

Create MCP selection component:
- Create new component in src/renderer/src/components/mcp-config/ (e.g.,
AgentMcpSelector.vue)
- Display available non-inmemory MCPs as checkboxes or toggles
- Accept agent ID as prop and load agent's current selections
- Emit selection changes to parent component

Integrate into agent settings:
- Locate agent configuration UI in src/renderer/src/components/ or settings
directory
- Embed MCP selection component in each agent's settings view
- Wire up save actions to call IPC methods that invoke
acpConfHelper.setAgentMcpSelections()

Add agent MCP status indicators:
- Add indicators in agent selection UI showing MCP count per agent
- Update agent cards/rows to display MCP count or status badge
- Add tooltips listing configured MCP names

💡 Iterate on the plan with: @coderabbitai <feedback>

Example Feedback
- @coderabbitai You can skip phase 3. Add a simple unit test case for phase 2.
- @coderabbitai For assumption 1 go ahead with option 3 and replan.

---

💬 Have feedback or questions? Drop into our discord or schedule a call!