The polite way to pipe Claude Max into any agent

What it does Meridian runs locally and translates requests from standard Anthropic or OpenAI API clients into calls to the official Claude Code SDK. Any coding agent that supports a custom base URL—OpenCode, Cline, Aider, Crush, Pi, or others—can route through it. Authentication, session management, streaming, and prompt caching all flow through the SDK’s documented query() function; Meridian just formats the output.

The interesting bit Rather than reverse-engineering or bypassing Anthropic’s platform controls, Meridian deliberately works inside them. It depends on the SDK for rate limiting, context compaction, and token refresh. The unusual part is an experimental toggle system that exposes SDK-exclusive features—auto-memory, dreaming, CLAUDE.md, extended thinking—to external agents that normally wouldn’t have access.

Key highlights

• Dual-protocol support: native Anthropic API plus OpenAI-compatible /v1/chat/completions for broader tool compatibility
• Two tool-execution modes: passthrough returns tool calls to the client, while internal mode lets the SDK handle execution directly
• Persistent sessions that survive proxy restarts, with concurrent parent and subagent requests and automatic model selection by context size
• Per-adapter experimental feature toggles configurable through a web UI, persisted to local config
• Built-in telemetry dashboard and Prometheus /metrics endpoint for token usage and cache efficiency
• Multi-profile support for switching between Claude accounts without restarting

Caveats

• SDK feature toggles, thinking passthrough, and related capabilities are explicitly marked experimental
• NixOS users must manually manage the OpenCode plugin path; the README warns that meridian setup writes absolute Nix store paths that break on rebuild

Verdict Worth trying if you pay for Claude Max but prefer a different agent frontend. Skip it if you already do everything inside Claude Code itself.