Making local models useful for coding: Hybrid cloud planning with local micro-patches
djpaul666 · reddit · 2026-08-17
The author built a "Local Coding Agent" to enable useful coding tasks on consumer GPUs (8GB VRAM) without relying on cloud credits.
How it works
- Architecture: Functions as an MCP server connected to primary cloud IDEs/agents via a delegatecode tool.
- Workflow: The cloud model handles high-level planning and isolates atomic tasks, handing off micro-patches to a local model (e.g., Gemma 4 2B/4B or Qwen3.8-27B).
Key Solutions
- Context Bloat: By receiving only isolated task context, the local model runs at full speed (60-85+ tok/s) without memory issues.
- Schema Failures: A deterministic rule-based engine corrects JSON/schema errors locally, avoiding wasteful cloud retries.
- Safety: Includes an isolated sandbox for testing and automatic git apply rollbacks if breaks occur.
More from coding & agent
- Why data readiness is the new bottleneck for enterprise AI agents — blaizedsouza · 2026-08-17
- Anthropic Leaks 4-Agent Setup That Cuts Codebase Audit from 3 Days to 20 Minutes — blaizedsouza · 2026-08-17
- Agent optimization: Replace intuition with A/B testing frameworks — blaizedsouza · 2026-08-17
- Building production-grade service layers for agents — blaizedsouza · 2026-08-17
- Multi-Tenancy Isolation Patterns for Agents: Memory, Policy, and Tool Security — blaizedsouza · 2026-08-17
- Reddit MCP Server: Zero-config integration and noise filtering for agents — Xabasis · 2026-08-17