Using prxy.monster with Claude Code

Claude Code speaks Anthropic Messages traffic, so it can point at prxy.monster with the Anthropic base URL. The value is the module pipeline around those calls: MCP pruning, cache hits, context preservation, and cost guards.

Setup time: 30 seconds.

Configure

export ANTHROPIC_BASE_URL=https://api.prxy.monster
export ANTHROPIC_API_KEY=prxy_live_xxxxxxxxxxxxxxxxxxxxxxxx

Restart Claude Code. That’s it.

Claude Code uses @anthropic-ai/sdk under the hood. It respects ANTHROPIC_BASE_URL automatically — no settings change, no config file edit, no restart of anything other than Claude Code itself.

Code change

None. Claude Code is a CLI; you don’t touch its code. Just the two env vars above.

Verify

prxy doctor

(or, if you don’t have the CLI:)

curl https://api.prxy.monster/health

Open Claude Code and watch its first few requests. They should succeed if routing and provider credentials are configured.

What you get

Concretely, on a typical Claude Code session:

Pain	Fix
MCP overhead — ~67k tokens loading every MCP definition before you type	`mcp-optimizer` — prunes irrelevant tools per query. ~67k → ~8k.
Rate limit at 20 minutes	`semantic-cache` returns repeats free. `exact-cache` catches identical re-asks.
Auto-compaction wipes context	`compaction-bridge` preserves the working set across compaction events. `rehydrator` restores the relevant context on the next message.
Sessions don’t compound — every Monday starts from zero	`patterns` logs successful fixes and re-injects them. The model gets faster on problem types you’ve solved before.
Long sessions hit context wall	`ipc` compresses old messages progressively (verbatim → truncated → summary → single sentence).

Recommended pipeline for Claude Code

This is the killer config:

PRXY_PIPE=mcp-optimizer,compaction-bridge,rehydrator,semantic-cache,patterns,ipc

Set it on the key with PATCH /v1/keys/:id using pipelineConfig, or set per-process in local mode:

export PRXY_PIPE=mcp-optimizer,compaction-bridge,rehydrator,semantic-cache,patterns,ipc

(Note: PRXY_PIPE is read by the proxy, not by Claude Code — for cloud, store the pipeline on the API key; for local, set it in the container environment.)

Streaming

Claude Code uses streaming heavily. prxy.monster passes through Anthropic’s per-block SSE format unmodified. Cache hits replay as synthetic SSE, indistinguishable to Claude Code’s parser.

Pricing reality check

You’re already paying Anthropic for tokens. prxy.monster doesn’t add per-token cost — you BYOK to Anthropic, and we charge a flat tier for the gateway.

Net effect for a typical Claude Code user:

Anthropic bill: same to lower when cache or optimization reduces provider work.
prxy.monster bill: $20/mo Pro, or Team for shared quota.
Your time: significantly more productive minutes per dollar.

Common issues

anthropic-beta headers — pass-through. Prompt caching, tool use, and other beta features all work.
MCP servers — Claude Code’s MCP wiring is unaffected. The mcp-optimizer module sees the tool definitions Claude Code sends and prunes irrelevant ones before the request hits Anthropic.
/clear — works the same. The patterns module preserves your learned patterns across /clear.
Local mode — point ANTHROPIC_BASE_URL=http://localhost:3099 to your local Docker container (see Local quickstart).

Worked example

Drop-in .env and a step-by-step setup walkthrough: github.com/Ekkos-Technologies-Inc/prxy-monster-examples/tree/main/examples/claude-code-setup