Integration: Custom MCP Client

Coverage status

MCP: ✅ — Any MCP-compatible client can connect via stdio or HTTP/SSE.

---

Setup (60-second path)

1. Install MemStrata and run memstrata init in your project.
2. Start the MCP server: memstrata mcp --stdio (stdio) or memstrata serve --port 8080 (HTTP).
3. Configure your client using the JSON snippet below.
4. Call context_query before sending prompts to your LLM.

---

What’s captured

• ✅ Any request that calls context_query before prompting — MemStrata compresses context and tracks token savings.
• ⚠️ If your client calls the LLM directly without going through context_query, those turns are not tracked.
• ❌ Responses from your LLM are not captured by default (the client handles those directly).

---

Overview

MemStrata exposes a standard MCP server that any MCP-compatible client can connect to. If your tool isn’t in the integration list, use this guide.

---

Transport options

MemStrata supports two MCP transports:

Transport	Command	When to use
stdio	memstrata mcp --stdio	Subprocess clients (most common)
HTTP/SSE	memstrata serve --port 8080	Network clients, multi-process setups

---

stdio transport (subprocess)

Most MCP clients launch the server as a subprocess and communicate over stdin/stdout. Configure your client to run:

memstrata mcp --stdio

Example MCP client configuration (JSON):

{
  "mcpServers": {
    "memstrata": {
      "command": "memstrata",
      "args": ["mcp", "--stdio"],
      "env": {}
    }
  }
}

On Windows, if memstrata is not on PATH:

{
  "mcpServers": {
    "memstrata": {
      "command": "C:\\Users\\you\\AppData\\Local\\Programs\\MemStrata\\memstrata.exe",
      "args": ["mcp", "--stdio"],
      "env": {}
    }
  }
}

---

HTTP/SSE transport

For clients that connect over HTTP or need a persistent server:

memstrata serve --port 8080

The MCP endpoint is at:

http://localhost:8080/mcp

SSE stream (for clients that use Server-Sent Events):

http://localhost:8080/mcp/sse

---

Available tools

MemStrata exposes three MCP tools:

context_query

Returns a compressed context slice for a given query. Call this before sending a prompt to your LLM.

Input schema:

{
  "type": "object",
  "properties": {
    "query": {
      "type": "string",
      "description": "Natural language query describing what context you need"
    },
    "max_tokens": {
      "type": "integer",
      "description": "Maximum tokens to return (default: 4000)",
      "default": 4000
    },
    "include_paths": {
      "type": "array",
      "items": { "type": "string" },
      "description": "Glob patterns to restrict context to specific paths (optional)"
    }
  },
  "required": ["query"]
}

Example call:

{
  "name": "context_query",
  "arguments": {
    "query": "How does authentication work in this codebase?",
    "max_tokens": 3000
  }
}

Response:

{
  "context": "..compressed code context slice..",
  "tokens_used": 2847,
  "tokens_saved": 4120,
  "files_included": ["src/auth/middleware.py", "src/auth/tokens.py"],
  "graph_depth": 2
}

---

index_status

Returns the current state of the local index.

Input: none required

Response:

{
  "indexed_files": 1247,
  "last_indexed": "2025-01-15T14:23:11Z",
  "coverage_percent": 98.2,
  "index_path": "/home/user/memstrata/index.db",
  "project_root": "/home/user/my-project"
}

---

savings_summary

Returns session token savings and estimated dollar amount.

Input: none required

Response:

{
  "session_tokens_saved": 84230,
  "session_dollar_savings": 0.42,
  "total_tokens_saved": 1240000,
  "total_dollar_savings": 6.20,
  "session_turns": 18
}

---

Python example

import subprocess, json

proc = subprocess.Popen(
    ["memstrata", "mcp", "--stdio"],
    stdin=subprocess.PIPE,
    stdout=subprocess.PIPE,
)

def call_tool(name, args):
    msg = json.dumps({"jsonrpc": "2.0", "id": 1, "method": "tools/call",
                      "params": {"name": name, "arguments": args}}) + "\n"
    proc.stdin.write(msg.encode())
    proc.stdin.flush()
    return json.loads(proc.stdout.readline())

# Initialize
proc.stdin.write(b'{"jsonrpc":"2.0","id":0,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"my-client","version":"1.0"}}}\n')
proc.stdin.flush()
proc.stdout.readline()  # read init response

# Query context
result = call_tool("context_query", {"query": "database schema and migrations"})
print(result["result"]["content"][0]["text"])

---

Verifying the connection

# Start the server
memstrata mcp --stdio &

# Send a JSON-RPC initialize message
echo '{"jsonrpc":"2.0","id":0,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1"}}}' | memstrata mcp --stdio

Expected response includes "serverInfo": {"name": "memstrata", "version": "5.x.x"}.

---

Troubleshooting

command not found: memstrata
Add MemStrata to your PATH, or use the full binary path in the command config.

Empty context response
Run memstrata status to verify the index is built. If indexed_files is 0, run memstrata index in your project directory first.

Tool not listed after initialize
Ensure you’re running MemStrata v5.0+. Run memstrata --version to check.

---

Related

• API Reference
• Quick Install
• Telemetry & Savings