$ head -n 1
Operational notes for running local coding agents: LM Studio as an OpenAI-compatible endpoint, task-specific presets, MCP scope, and context budgets that keep edit loops responsive.
$ grep -i "problem"
Local inference is most useful when the runtime is treated as a small service with explicit constraints. Each model has different latency, memory pressure, context length, tool-call behavior, and failure modes, so the routing policy matters as much as the model list.
For coding-agent work, the default goal is a short feedback loop: inspect a bounded surface, produce a patch, run verification, and summarize state before the next phase.
$ grep -i "runtime shape"
LM Studio provides the OpenAI-compatible endpoint for local clients. Presets carry practical defaults such as context length, temperature, GPU memory expectations, and when to stop or split work before overflow.
Coding clients can route shallow edits to fast local coder models, planning passes to larger local reasoning models, and only escalate to hosted models when latency, accuracy, or context headroom changes the result.
$ grep -i "practical rules"
Keep one active local model for stable edit loops; running several large contexts concurrently can turn memory pressure into latency spikes or unload/reload churn. Use low-temperature settings for patches and reserve broader sampling for exploration.
Treat MCP tools as capability grants. Enable the minimum useful set for the task, avoid secret-bearing outputs in tool transcripts, and prefer small verifiable phases over letting a single prompt accumulate the whole repo.