mddlmn

A local control layer for Claude Code and Anthropic API traffic. Inspect, edit, redact, inject, and gate every request before it reaches the model.

What is mddlmn?

mddlmn (middleman) is a local-only proxy that sits between an Anthropic-API-compatible client — Claude Code, the Anthropic SDK, your own agent — and api.anthropic.com. It captures every request and response, classifies what's inside, and gives you a side-panel UI to inspect, edit, hold, redact, or inject traffic before it goes upstream.

This extension is the VS Code shell. On activation it:

• Starts the local mddlmn proxy on an available port
• Injects ANTHROPIC_BASE_URL into Claude Code's settings and your integrated terminals — no manual env-var setup
• Opens the inspector in a webview panel
• Shows the active proxy port in the status bar

Your API key is never read or stored — auth headers pass through unchanged.

Features

See everything

• Per-section classification: every captured turn is broken into system prompt, tool definitions, user text, injected context, assistant text, tool-use, tool-result, and extended thinking blocks
• Token counts per section, per turn, per session
• Live updates over WebSocket as new requests land
• Diff any two requests side-by-side to see what changed
• Timeline view of a whole session

Change everything

• Gate: arm the gate to hold every outgoing request until you approve, edit, or abort it. The agent loop blocks cleanly on the network rather than spinning.
• Section editing: full in-UI editor for held requests — modify text, drag-reorder content blocks and messages, delete (with restore), edit JSON tool definitions, swap models from a dropdown
• Secret redaction: built-in patterns strip Anthropic / OpenAI / AWS keys and PEM private key blocks before the request is shown or forwarded. Add your own rules at any time.
• Prompt injection: append standing context to the system prompt, prepend to the last user message, or insert a new user turn — scoped to top-level conversations, tool-chain steps, or both
• Memory injection: a persistent store of context snippets auto-injected into every request. Each entry is always on, session-only (gone on restart), or conditional (injected when a regex matches the last user message), with optional expiry.
• Settings tab: manage redaction rules, injection rules, and memory entries from one panel
• Synthetic abort: cancelling a held request returns a clean SSE end_turn so the agent loop completes instead of retrying

Honest about what's what

Every held request shows a TOP / TOOL / AUX badge and the model name in the header — so when Claude Code fires a haiku summarization probe alongside your main opus generation, you know which one you're editing.

Quick start

1. Install the extension and reload VS Code
2. The proxy starts automatically; ANTHROPIC_BASE_URL is injected into Claude Code and new terminals
3. Use Claude Code or any Anthropic-API-compatible client normally — every request is captured
4. Open the panel: Command Palette → "mddlmn: Open Panel"
5. Click GATE in the header to start intercepting requests

How it works

agent → mddlmn proxy → api.anthropic.com
        │
        ├─ inbound middleware  (redaction)
        ├─ canonical conversation
        ├─ gate (hold for approval / edits)
        └─ outbound middleware (injection, memory)

The proxy maintains its own canonical copy of the conversation, so edits and aborts persist across the client's stateless replays. Cache-control markers are normalized so requests never exceed Anthropic's 4-block cache limit. Aborts return a synthetic SSE end_turn so the agent loop completes naturally.

Where data lives

Everything stays on your machine, at ~/.mddlmn/:

Override the root with the MDDLMN_CONFIG_DIR environment variable. Session-scoped memory entries are never written to disk — they live only in RAM and vanish on proxy restart.

Privacy

• 100% local. No telemetry. No cloud calls except to the upstream Anthropic API on your behalf.
• Your API key is passed through verbatim — mddlmn never reads or stores it.
• Captured prompts, tool outputs, and model responses are written to local files. Treat them as sensitive.

Source

Open source: github.com/juliank1m/mddlmn

Issues and feature requests welcome.