Chkim Su Forge Editor MCP Daemon Isolation

name: mcp-daemon-isolation description: Context isolation for query-type MCP tools (LSP, search, database) via external CLI. Use when MCP query results consume too many context tokens. allowed-tools: ["Read", "Bash", "Grep", "Glob"]

MCP Daemon Isolation for Query-Type MCPs

External CLI pattern for isolating query-type MCP tool results from main context.

Query-Type MCP Definition

Identification criteria: Tools with find_*, search_*, get_*, list_* patterns

Problem

Claude Session
├── mcp__serena__find_symbol("UserService")
│   └── Result: 2,500 tokens (full JSON)    ← CONSUMED
└── Context budget: rapidly depleting

Daemon isolates tool definitions (~350 tokens × tools), but results still consume context.

Solution: External CLI + Structured Extraction

Claude Session
├── Bash: serena-query find_symbol UserService
│   └── stdout: "• UserService [Class] @ src/user.py:42"
│       (~50 tokens)                      ← MINIMAL
└── Full result stored: /tmp/serena-result.json

Key insight: Claude IS the LLM. Structured extraction provides 95%+ token savings.

Architecture

┌──────────────┐     ┌──────────────────┐     ┌──────────────┐
│ Claude       │────▶│ serena-query     │────▶│ Serena       │
│ Session      │     │ (external CLI)   │     │ Daemon :8765 │
└──────────────┘     └──────────────────┘     └──────────────┘
     Bash (~50 tok)     SSE + MCP Protocol       29 tools

For protocol details: Read("references/mcp-sse-protocol.md")

Structured Extractors

For formatter code: Read("references/extractors-and-examples.md")

Usage

Output Modes

serena-query find_symbol UserService                    # summary (default)
serena-query find_symbol UserService --mode location    # location only (for Read)
serena-query find_symbol UserService --mode full        # full JSON

Basic Commands

serena-query list_dir .
serena-query get_symbols_overview src/main.py --depth 1
serena-query find_symbol UserService --path src/
serena-query search_for_pattern "class.*Service" --path src/
serena-query find_symbol UserService --output /tmp/result.json

Optimal Workflow: Explore-Locate-Read

1. Explore (--mode summary)   ~25 tokens  "What symbols exist?"
2. Locate (--mode location)   ~15 tokens  "Where exactly is it?"
3. Read (Read tool)           actual size  Only needed code

For detailed examples: Read("references/extractors-and-examples.md")

Installation

pip install httpx
cp scripts/serena-query ~/.local/bin/
chmod +x ~/.local/bin/serena-query

Daemon Setup

uvx --from git+https://github.com/oraios/serena \
  serena start-mcp-server --transport sse --port 8765 --project-from-cwd

Decision: Structured vs LLM Summarization

Claude IS the LLM consuming the output - no need for additional summarization.

Isolation Strategy Guide

Common Issues

References

Implementation

• serena-query CLI
• Implementation Details
• v2 Design

Protocol & Examples

• MCP-SSE Protocol
• Extractors & Examples
• Integration Guide

Related

• MCP Gateway Patterns

Converted and distributed by TomeVault — claim your Tome and manage your conversions.