💡 Per-Tier MCP Tool Gating with Selective LSP Tool Exposure

[github-actions[bot]]

Summary

Extend engine.sh's existing per-tier model selection to include per-tier MCP tool sets. Triage gets zero MCP tools (speed), deep gets navigation tools (definition/references/diagnostics), audit gets full LSP + speculative execution (preview_edit, simulate_chain for blast-radius analysis). This aligns MCP tool overhead with tier budgets and model capabilities — Haiku/Flash don't benefit from 65 tools, but Opus/Fable do.

Market Signal

MindStudio's analysis (2026) quantifies MCP tool overhead at ~200 tokens per tool definition per turn, with 15–20K tokens consumed by mid-range tool sets before any actual query. The recommended mitigation is "focused agents with minimal, relevant tool sets." GitHub Copilot CLI's LSP Setup skill (June 10, 2026) takes a similar selective approach, exposing only the tools the current task needs rather than the full LSP surface.

User Signal

Discussion #578's phased rollout recommendation already suggests gating LSP to deep/audit. engine.sh already has per-tier model chains (CLAUDE_DEEP_MODEL_CHAIN, CLAUDE_AUDIT_MODEL_CHAIN), per-tier timeouts (TRIAGE_TIMEOUT_SEC=300, DEEP_TIMEOUT_SEC=600), and per-tier allowed-tools (--allowed-tools "Bash,Read,Grep,Glob") — the architecture is pre-wired for per-tier MCP configs. The Token Cost Observatory would capture the per-tier token delta.

Technical Opportunity

engine.sh's run_agentic() and run_triage() already distinguish tiers via the $tier parameter. The proposed architecture:

• mcp-config-deep.json — 8 navigation tools: definition, references, diagnostics, hover, documentSymbol, workspace_symbol, type_definition, declaration
• mcp-config-audit.json — deep set + 6 advanced tools: preview_edit, simulate_chain, rename, call_hierarchy, type_hierarchy, code_action
• Triage — no MCP config (stays MCP-free for speed)

The --mcp-config flag on the claude CLI supports multiple configs that merge, so tier-specific configs compose with any future non-LSP MCP servers (e.g., SonarQube Agentic Analysis).

Token budget per tier:

Tier	Model	MCP Tools	Overhead/turn	Budget	Overhead %
Triage	Haiku 4.5	0	0	300s	0%
Deep	Opus 4.8	8	~1,600 tok	600s	<1% of 200K window
Audit	Fable 5	14	~2,800 tok	600s	~1.4% of 200K window

Assessment

Dimension	Score	Rationale
Feasibility	high	Extends existing per-tier architecture with JSON config files + one --mcp-config flag
Impact	med	Enables cost-controlled LSP adoption; prevents triage regression
Urgency	high	Must be in place before LSP pilot to prevent token waste on triage tier

Adversarial Review

Strongest objection: Per-tier MCP configs add operational complexity — more files to maintain, more cache keys, more failure modes. The Gemini and Copilot engines may not support the same MCP config format, creating engine-specific branching.

Rebuttal: Complexity is bounded: 2–3 JSON files, managed alongside the existing per-tier model chains. Engine branching already exists in set_engine_config() — MCP config follows the same pattern. For non-Claude engines, the MCP config is simply skipped (graceful degradation, exactly as today). The alternative — exposing all tools to all tiers — wastes tokens on triage and overwhelms lightweight models with tool definitions they can't effectively use.

Suggested Next Step

Define the tool subsets per tier (using agent-lsp's tool names). Create the JSON config files under scripts/mcp-configs/. Wire into engine.sh's run_agentic() with a conditional --mcp-config flag gated on $tier and $REVIEW_ENGINE=claude.

_{🤖 Proposed by Mary (BMAD Strategic Business Analyst) · companion to Discussion #578}