The Token Problem

Loading all MCP tool schemas upfront burns context window

❌ Today: Upfront Loading
~50
tool schemas in system prompt
Unused schemas
Actually called
Every turn pays for schemas the agent never uses. Most conversations touch 2-3 tools max.
✅ Better: Lazy Discovery
0
schemas loaded at start
Tokens saved
On-demand load
Agent searches → discovers matching tool → loads only what it needs. Context stays clean.

Proposed Architecture

Clippy → ODR → MCP Server (no intermediary CLI needed)

Step 1
Route
odr mcp search "…"
Find the right server
by keyword match
Step 2
Launch
odr mcp --proxy <server>
Sandboxed stdio pipe
with IT controls
Step 3
Call
Native MCP via SDK
Clippy speaks MCP
directly — no wrapper
🔍
Lazy Discovery
No upfront schema loading
🛡️
Sandboxed
ODR isolates server processes
📋
Audit Trail
IT visibility into tool calls
⚡
No Extra Hop
SDK → MCP server directly

Platform Strategy

Best tool for each platform — no one-size-fits-all

Windows
ODR + Native SDK
  • ✓ Discovery via odr mcp search
  • ✓ Sandboxed server lifecycle
  • ✓ IT admin policy controls
  • ✓ WAM-based auth (Entra ID)
  • ✓ Audit logging built-in
  • ✓ Native MCP through Clippy SDK
Mac / Linux
mcp-cli (Go)
  • ✓ Discovery via local config
  • ✓ Stdio + SSE transports
  • ✓ Stateful session persistence
  • ✓ Custom OAuth flow
  • ✓ Cross-platform Go binary
  • ✓ Self-contained, no OS dependency
Key insight: Both paths solve the same token problem — lazy discovery instead of upfront loading. ODR provides managed infrastructure on Windows; mcp-cli provides the portable fallback everywhere else.