Claude builds for the cloud by default.
A new skill that teaches Claude how to build local agents.
If you’ve tried using Claude to build local AI agents, you know the cycle. It treats local endpoints like they have infinite cloud context, ignores VRAM limits, cuts off generation early, and makes the exact same setup mistakes over and over again.
We got tired of fixing the same harness bugs, so we built local-agents: a skill that teaches Claude the actual mechanics of running agents on local hardware.
Install in Claude Code
/plugin marketplace add ghosthat-studio/marketplace
/plugin install local-agents@ghosthat-studioWhy we built it: read the write-up.
The problem: Claude’s cloud habits
When Claude writes an agent loop for an Ollama, vLLM or llama.cpp endpoint, it assumes cloud-scale infrastructure. That leads to predictable frustration.
It misreads stop reasons
Claude defaults to fixed iteration caps or times out early, treating a hit context window or thinking loop as a completed task instead of continuing execution.
It ignores VRAM constraints
It doesn’t check whether weights, KV cache and context window actually fit in memory, so the model quietly falls back to slow CPU inference.
It overflows context
Cloud habits mean uncalibrated tool schemas and full history dumps, filling up local context windows before the model even finishes reasoning.
You shouldn’t have to correct Claude on local memory, engine quirks and context budgeting every time you start a new project.
What changes with the skill
When this skill is loaded, Claude stops guessing and applies a tested, step-by-step workflow for local hardware.
Measures the stack first
Before writing code, Claude runs a probe to detect actual VRAM, context limits, device placement (GPU or CPU) and running services.
Handles stop reasons correctly
Loops end on explicit stop reasons (
tool_calls,stop,length), keeping output budgets intact and generation timeouts off the model.Budgets context and rolls history
System prompts, tool definitions, thinking reserves and answers are budgeted per turn. History rolls cleanly before anything is truncated.
Guards execution, not generation
Reads run freely, but state-changing actions are guarded by consequence, which gives local models the freedom to act safely.
How to use it
Claude Code
/plugin marketplace add ghosthat-studio/marketplace
/plugin install local-agents@ghosthat-studioClaude.ai, desktop and API
- Download the skill.
- Upload it under Customize → Skills.
Once installed, Claude automatically applies these practices whenever you build, adjust or debug a local agent loop.
Every install option is in the guide on GitHub.
Rather have us build it for you?
We design and build custom agents that run on your own hardware. Tell us what you want done.