|
| 1 | +# Feature Specification: MCP Direct Endpoint |
| 2 | + |
| 3 | +**Feature Branch**: `025-mcp-direct-endpoint` |
| 4 | +**Created**: 2026-01-26 |
| 5 | +**Status**: Draft |
| 6 | +**Input**: User description: "Add MCP direct endpoint to expose all proxied tools directly alongside internal tools, with server:tool namespacing for upstream tools" |
| 7 | +**Related Issue**: https://github.com/smart-mcp-proxy/mcpproxy-go/issues/279 |
| 8 | + |
| 9 | +## Problem Statement |
| 10 | + |
| 11 | +Currently, MCPProxy exposes tools through a search-first workflow: |
| 12 | +1. Clients call `retrieve_tools` to search for tools |
| 13 | +2. Clients call `call_tool_read/write/destructive` with `server:tool` format |
| 14 | + |
| 15 | +This workflow provides significant token savings (~99%) for large tool sets. However, some clients and use cases need direct access to all tools without the search intermediary, especially as AI platforms adopt built-in tool search capabilities (per Anthropic's "Advanced Tool Use" announcement). |
| 16 | + |
| 17 | +## User Scenarios & Testing *(mandatory)* |
| 18 | + |
| 19 | +### User Story 1 - Direct Tool Access (Priority: P1) |
| 20 | + |
| 21 | +As an MCP client user, I want to connect to a single endpoint that exposes all available tools directly, so I can use any tool without first searching for it. |
| 22 | + |
| 23 | +**Why this priority**: This is the core value proposition - enabling direct tool access for clients that don't need the search-first workflow. |
| 24 | + |
| 25 | +**Independent Test**: Can be fully tested by connecting an MCP client to the `/mcp/direct` endpoint and verifying all upstream tools appear in the tools/list response alongside internal tools. |
| 26 | + |
| 27 | +**Acceptance Scenarios**: |
| 28 | + |
| 29 | +1. **Given** the direct endpoint is enabled and upstream servers are connected, **When** a client requests the tools list from `/mcp/direct`, **Then** the response includes all tools from connected upstream servers with `server:tool` naming format. |
| 30 | + |
| 31 | +2. **Given** the direct endpoint is enabled, **When** a client requests the tools list, **Then** the response includes all internal MCPProxy tools (retrieve_tools, call_tool_*, upstream_servers, etc.) without a server prefix. |
| 32 | + |
| 33 | +3. **Given** the direct endpoint is enabled and a tool exists as `github:create_issue`, **When** a client calls this tool directly, **Then** the call is routed to the github upstream server and executed successfully. |
| 34 | + |
| 35 | +--- |
| 36 | + |
| 37 | +### User Story 2 - Configuration Toggle (Priority: P1) |
| 38 | + |
| 39 | +As a system administrator, I want to enable or disable the direct endpoint via configuration, so I can choose whether to expose all tools directly or require the search-first workflow. |
| 40 | + |
| 41 | +**Why this priority**: Essential for deployment flexibility - some environments want search-first (token savings), others want direct access. |
| 42 | + |
| 43 | +**Independent Test**: Can be tested by toggling the configuration flag and verifying the endpoint availability changes accordingly. |
| 44 | + |
| 45 | +**Acceptance Scenarios**: |
| 46 | + |
| 47 | +1. **Given** the configuration option `enable_direct_endpoint` is set to `true`, **When** the server starts, **Then** the `/mcp/direct` endpoint is available and functional. |
| 48 | + |
| 49 | +2. **Given** the configuration option `enable_direct_endpoint` is set to `false` (or not set), **When** a client attempts to access `/mcp/direct`, **Then** the endpoint is not available (returns 404). |
| 50 | + |
| 51 | +3. **Given** the direct endpoint is enabled, **When** viewing server status or logs, **Then** the system indicates the direct endpoint is active. |
| 52 | + |
| 53 | +--- |
| 54 | + |
| 55 | +### User Story 3 - Security Exclusions (Priority: P2) |
| 56 | + |
| 57 | +As a security-conscious administrator, I want quarantined servers' tools excluded from the direct endpoint, so I don't accidentally expose unapproved tools to clients. |
| 58 | + |
| 59 | +**Why this priority**: Important for security, but secondary to core functionality since quarantine is an existing security feature. |
| 60 | + |
| 61 | +**Independent Test**: Can be tested by quarantining a server and verifying its tools disappear from the direct endpoint's tool list. |
| 62 | + |
| 63 | +**Acceptance Scenarios**: |
| 64 | + |
| 65 | +1. **Given** a server is quarantined, **When** a client requests the tools list from `/mcp/direct`, **Then** tools from the quarantined server are not included in the response. |
| 66 | + |
| 67 | +2. **Given** a server is disabled, **When** a client requests the tools list from `/mcp/direct`, **Then** tools from the disabled server are not included in the response. |
| 68 | + |
| 69 | +3. **Given** a server is disconnected, **When** a client requests the tools list from `/mcp/direct`, **Then** tools from the disconnected server are not included in the response. |
| 70 | + |
| 71 | +--- |
| 72 | + |
| 73 | +### User Story 4 - Dynamic Tool Synchronization (Priority: P2) |
| 74 | + |
| 75 | +As an MCP client user, I want the direct endpoint's tool list to update automatically when upstream servers connect or disconnect, so I always see the current available tools. |
| 76 | + |
| 77 | +**Why this priority**: Important for reliability but secondary to initial tool exposure. |
| 78 | + |
| 79 | +**Independent Test**: Can be tested by connecting/disconnecting an upstream server and observing the tool list update on the direct endpoint. |
| 80 | + |
| 81 | +**Acceptance Scenarios**: |
| 82 | + |
| 83 | +1. **Given** the direct endpoint is active and an upstream server connects, **When** the server's tools become available, **Then** the direct endpoint's tool list is updated to include the new tools. |
| 84 | + |
| 85 | +2. **Given** the direct endpoint is active and an upstream server disconnects, **When** the server becomes unavailable, **Then** the direct endpoint's tool list is updated to remove the server's tools. |
| 86 | + |
| 87 | +3. **Given** the direct endpoint is active, **When** an upstream server's tools change (via tools/list_changed notification), **Then** the direct endpoint's tool list reflects the changes. |
| 88 | + |
| 89 | +--- |
| 90 | + |
| 91 | +### User Story 5 - Tool Annotation Preservation (Priority: P3) |
| 92 | + |
| 93 | +As an MCP client developer, I want upstream tool annotations (readOnlyHint, destructiveHint, etc.) preserved on the direct endpoint, so my client can make informed decisions about tool behavior. |
| 94 | + |
| 95 | +**Why this priority**: Nice-to-have for client intelligence but not critical for basic functionality. |
| 96 | + |
| 97 | +**Independent Test**: Can be tested by verifying tool annotations from upstream servers appear correctly on tools retrieved from the direct endpoint. |
| 98 | + |
| 99 | +**Acceptance Scenarios**: |
| 100 | + |
| 101 | +1. **Given** an upstream tool has `readOnlyHint: true` annotation, **When** the tool is listed on the direct endpoint, **Then** the annotation is preserved in the tool definition. |
| 102 | + |
| 103 | +2. **Given** an upstream tool has `destructiveHint: true` annotation, **When** the tool is listed on the direct endpoint, **Then** the annotation is preserved in the tool definition. |
| 104 | + |
| 105 | +--- |
| 106 | + |
| 107 | +### Edge Cases |
| 108 | + |
| 109 | +- What happens when all upstream servers are disabled or quarantined? → Only internal tools are listed |
| 110 | +- What happens when two upstream servers have tools with the same name? → They're disambiguated by server prefix (e.g., `serverA:read_file` vs `serverB:read_file`) |
| 111 | +- What happens if an upstream server provides a tool named like an internal tool? → Server-prefixed name prevents collision (e.g., `myserver:retrieve_tools` vs `retrieve_tools`) |
| 112 | +- How does the system handle rapid connect/disconnect cycles? → Tool list updates atomically on each state change |
| 113 | + |
| 114 | +## Requirements *(mandatory)* |
| 115 | + |
| 116 | +### Functional Requirements |
| 117 | + |
| 118 | +- **FR-001**: System MUST provide a new MCP endpoint at `/mcp/direct` when the feature is enabled |
| 119 | +- **FR-002**: System MUST expose all tools from connected, enabled, non-quarantined upstream servers on the direct endpoint |
| 120 | +- **FR-003**: System MUST expose all internal MCPProxy tools (retrieve_tools, call_tool_read, call_tool_write, call_tool_destructive, upstream_servers, quarantine_security, read_cache, code_execution, list_registries, search_servers) on the direct endpoint |
| 121 | +- **FR-004**: System MUST namespace upstream tools using `server:tool` format to prevent collisions |
| 122 | +- **FR-005**: System MUST NOT namespace internal tools (they appear without prefix) |
| 123 | +- **FR-006**: System MUST exclude tools from quarantined servers |
| 124 | +- **FR-007**: System MUST exclude tools from disabled servers |
| 125 | +- **FR-008**: System MUST exclude tools from disconnected servers |
| 126 | +- **FR-009**: System MUST synchronize the tool list when upstream server state changes (connect, disconnect, quarantine, enable/disable) |
| 127 | +- **FR-010**: System MUST preserve tool annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint, title) from upstream servers |
| 128 | +- **FR-011**: System MUST route tool calls to the appropriate upstream server based on the server prefix |
| 129 | +- **FR-012**: System MUST handle internal tool calls by delegating to existing handler logic |
| 130 | +- **FR-013**: System MUST provide a configuration option `enable_direct_endpoint` to enable/disable this feature |
| 131 | +- **FR-014**: System MUST default the `enable_direct_endpoint` configuration to `false` (opt-in) |
| 132 | +- **FR-015**: System MUST return 404 for `/mcp/direct` when the feature is disabled |
| 133 | + |
| 134 | +### Key Entities |
| 135 | + |
| 136 | +- **DirectMCPServer**: A separate MCP server instance that manages the direct endpoint, distinct from the search-based MCPProxyServer |
| 137 | +- **Tool Namespace**: The `server:tool` naming convention that maps upstream tools to their source server |
| 138 | +- **Tool Synchronization**: The mechanism that keeps the direct endpoint's tool list in sync with upstream server state |
| 139 | + |
| 140 | +## Success Criteria *(mandatory)* |
| 141 | + |
| 142 | +### Measurable Outcomes |
| 143 | + |
| 144 | +- **SC-001**: Clients connecting to `/mcp/direct` receive a complete tool list within 500ms of connection |
| 145 | +- **SC-002**: Tool list updates reflect upstream server changes within 2 seconds of state change |
| 146 | +- **SC-003**: 100% of upstream tool calls through the direct endpoint succeed when the upstream server is healthy |
| 147 | +- **SC-004**: Tool annotations are preserved with 100% fidelity from upstream to direct endpoint |
| 148 | +- **SC-005**: Zero tools from quarantined/disabled/disconnected servers appear in the direct endpoint's tool list |
| 149 | + |
| 150 | +## Assumptions |
| 151 | + |
| 152 | +- The existing `EventTypeServersChanged` event provides sufficient notification for tool synchronization |
| 153 | +- The `StateView` contains cached tool information suitable for rebuilding the tool list |
| 154 | +- The `mcp-go` library's `SetTools()` method provides atomic tool list replacement |
| 155 | +- Internal tool handlers can be shared between the existing MCPProxyServer and the new DirectMCPServer |
| 156 | + |
| 157 | +## Out of Scope |
| 158 | + |
| 159 | +- Changing the behavior of the existing `/mcp` endpoint |
| 160 | +- Per-server or per-tool filtering on the direct endpoint (all-or-nothing exposure) |
| 161 | +- Session sharing between `/mcp` and `/mcp/direct` endpoints |
| 162 | +- WebUI integration for the direct endpoint |
| 163 | + |
| 164 | +## Commit Message Conventions *(mandatory)* |
| 165 | + |
| 166 | +When committing changes for this feature, follow these guidelines: |
| 167 | + |
| 168 | +### Issue References |
| 169 | +- ✅ **Use**: `Related #279` - Links the commit to the issue without auto-closing |
| 170 | +- ❌ **Do NOT use**: `Fixes #279`, `Closes #279`, `Resolves #279` - These auto-close issues on merge |
| 171 | + |
| 172 | +**Rationale**: Issues should only be closed manually after verification and testing in production, not automatically on merge. |
| 173 | + |
| 174 | +### Co-Authorship |
| 175 | +- ❌ **Do NOT include**: `Co-Authored-By: Claude <noreply@anthropic.com>` |
| 176 | +- ❌ **Do NOT include**: "🤖 Generated with [Claude Code](https://claude.com/claude-code)" |
| 177 | + |
| 178 | +**Rationale**: Commit authorship should reflect the human contributors, not the AI tools used. |
| 179 | + |
| 180 | +### Example Commit Message |
| 181 | +``` |
| 182 | +feat: add MCP direct endpoint for all-tools exposure |
| 183 | +
|
| 184 | +Related #279 |
| 185 | +
|
| 186 | +Adds a new /mcp/direct endpoint that exposes all upstream tools directly |
| 187 | +alongside internal MCPProxy tools, enabling clients to use tools without |
| 188 | +the search-first workflow. |
| 189 | +
|
| 190 | +## Changes |
| 191 | +- Add EnableDirectEndpoint config option (default: false) |
| 192 | +- Create DirectMCPServer with tool synchronization |
| 193 | +- Register internal tools and upstream tools with server:tool namespacing |
| 194 | +- Mount endpoint at /mcp/direct when enabled |
| 195 | +
|
| 196 | +## Testing |
| 197 | +- E2E tests for tool listing and calling |
| 198 | +- Unit tests for tool synchronization logic |
| 199 | +``` |
0 commit comments