DC-KALIC Parity Breakthrough

The Problem: Asymmetric Capabilities

DC (Desktop Claude) runs in claude.ai with access to 119+ MCP tools. KALIC runs in WSL via Claude Code with native Linux tools but zero MCP access. Two agents in the same ecosystem, speaking different languages.

The asymmetry wasn't just inconvenient—it was crippling collaboration. DC could instant-search 2 million files via Everything MCP in milliseconds. KALIC had to wait 10+ seconds for find to crawl directories. DC could capture screenshots. KALIC was blind to the Windows GUI entirely.

The Audit: Identifying Gaps

DC sent KALIC a capability inventory via the file messaging system. KALIC's response was surgical:

Three critical gaps emerged: instant file search, screenshot capture, and context tracking. The mission was clear.

The Solution Sprint: 24 Hours, 5 Tools

The Innovation: Dual-Mode MCP Servers

// Rust pseudo-code for dual-mode detection
fn main() {
    let args: Vec<String> = env::args().collect();
    
    if args.len() > 1 {
        // CLI mode: parse subcommands
        match args[1].as_str() {
            "search" => cli_search(&args[2..]),
            "ext" => cli_ext(&args[2..]),
            "recent" => cli_recent(&args[2..]),
            "status" => cli_status(),
            _ => print_help()
        }
    } else {
        // MCP mode: stdio JSON-RPC
        run_mcp_server()
    }
}

The everything-mcp-rs binary now serves both masters:

# MCP mode (Claude Desktop invokes this)
$ everything-mcp-rs.exe
{"jsonrpc":"2.0","method":"initialize"...}

# CLI mode (KALIC or any human can use this)
$ everything-mcp-rs.exe search "carbyne" -n 20
Found 147 items (showing 20):
[FILE] X:\carbyne-cli\target\release\carbyne.exe
[DIR] C:\Dev\carbyne
...

Why This Matters

1. Universal Tool Access

Any tool built as dual-mode is accessible to ANY agent, regardless of whether they have MCP support. WSL agents, SSH sessions, cron jobs, humans in terminals—everyone gets the same capabilities.

2. Testing Without Infrastructure

MCP servers are notoriously hard to test. You need Claude Desktop or a mock MCP client. CLI mode lets you test functionality directly: ./my-mcp search "test" -n 5. If the CLI works, the MCP works.

3. Graceful Degradation

MCP connection drops? Timeout? Fall back to CLI. The capability remains available even when the protocol fails.

4. Composability

CLI tools pipe. everything-mcp-rs search "*.rs" | grep "carbyne" works. MCP doesn't compose—it's request/response only.

The Complete Toolbox

After 24 hours, KALIC's capability gaps are closed. The unified toolbox at C:\PhiDEX\tools\:

The Communication Layer

None of this would work without reliable agent-to-agent messaging. The msgqueue system:

C:\PhiDEX\msgqueue\
├── dc\
│   ├── inbox\       ← KALIC writes here
│   └── processed\   ← DC moves read messages
└── kalic\
    ├── inbox\       ← DC writes here
    └── processed\   ← KALIC moves read messages

Filename: from_<SENDER>_YYYYMMDD_HHMMSS.txt
Signal: "$" character = "check your inbox"

The $ bump protocol: When DC writes a message to KALIC's inbox, it sends a single $ character to KALIC's terminal via AHK automation. KALIC sees $, checks inbox, reads message. No polling. No websockets. Just files and a signal.

Lessons Learned

1. Parity Enables Collaboration

Asymmetric capabilities create bottlenecks. When both agents can do the same things, work flows naturally.

2. CLI is Universal

Protocols come and go. Command-line interfaces work everywhere, forever. Build CLI first, add protocol adapters second.

3. Dual-Mode is the Pattern

Every MCP server should detect its invocation context and adapt. One binary, multiple access paths.

4. Check Before Building

KALIC now runs phi-find <topic> before creating anything. If it exists, use it. If not, build it and document it.

5. File Messaging Works

Fancy IPC is fragile. Files are robust. The msgqueue system has exchanged 50+ messages with zero failures.

What's Next

The dual-mode pattern will propagate to other MCP servers:

• git-mcp-rs → CLI subcommands for common operations
• screen-mcp-rs → CLI for screenshots without MCP overhead
• sigcon6 → Network diagnostics from any terminal

The goal: every tool accessible everywhere. No agent left behind.