Coding assistant

For Claude Code, Cline, Continue.dev, custom MCP-based coding agents — anything that hits the gateway with a fat tool list and runs for hours per session.

What this pipeline is good at

Stripping the 60-tool MCP catalog down to the 4 the user is actually asking about.
Keeping context budget healthy across hours of back-and-forth.
Learning your codebase’s quirks, your team’s conventions, your common gotchas.
(v1.1) Surviving Claude Code’s auto-compaction.

The pipeline

Env var


PRXY_PIPE='mcp-optimizer,patterns,ipc'

When v1.1 ships, add prompt-optimizer after ipc and compaction-bridge at the end:


PRXY_PIPE='mcp-optimizer,prompt-optimizer,patterns,ipc,rehydrator,compaction-bridge'

Why this order

mcp-optimizer first — strips tools before anything else looks at the prompt. The kept set is stable per session, so prompt cache prefix doesn’t shatter.
patterns — injects relevant past fixes BEFORE the prompt is finalized. Pattern injection adds tokens, so it goes after token-stripping.
ipc last — measures the final prompt size after everything else has touched it. Compresses oldest turns to fit under target.

What you’ll see

After a week of using this pipeline against the same codebase:

First-day metrics: patterns.injected ~0, mcp-optimizer.tokens.saved ~50k/request, ipc kicks in around hour 2.
Week-2 metrics: patterns.injected 3-4 per request, response quality on recurring problems noticeably faster.
Month-2 metrics: model “knows” your codebase. The patterns table in ~/.prxy/prxy.db (local) or your Postgres has hundreds of forged patterns. Repeated questions get answered in fewer turns.

Variants

Heavy MCP user (Claude Code with many connected servers):


pipeline:
  - mcp-optimizer:
      relevanceThreshold: 0.5    # less aggressive — keep more tools
      preserveTools: [read_file, write_file, bash, grep, glob, edit_file]
      minTools: 3
  - patterns: { maxInjected: 5 }
  - ipc: { targetUtilization: 0.7 }

Pattern-focused (you trust your codebase patterns more than tool relevance):


pipeline:
  - patterns:
      maxInjected: 8              # inject more past fixes
      minSuccessRate: 0.5         # less conservative
  - mcp-optimizer:
      relevanceThreshold: 0.65
  - ipc: { targetUtilization: 0.75 }