skills.rest
dxiiren
dxiirenCOMMUNITY
Explore all skills by this author →
GitHub

setup-claude-local

Installs and verifies the claude-local launcher running Claude Code against a self-hosted vLLM server.

Updated Aug 1, 2026
Search Tags:#claude code#setup#ollama#powershell#vllm#self-hosted llm#shim proxy
One-click install
npx skills add https://github.com/dxiiren/project-skeleton --skill setup-claude-local-dxiiren
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: setup-claude-local
Source: https://github.com/dxiiren/project-skeleton/tree/main/.claude/skills/setup-claude-local
Command: npx skills add https://github.com/dxiiren/project-skeleton --skill setup-claude-local-dxiiren

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Running Claude Code against a self-hosted vLLM model requires a local shim, per-machine configuration, and careful verification — misconfigured setups fail with cryptic 400 errors, truncated prompts, or silent context clamping. This Skill installs the claude-local launcher, proves it works with a live print-mode run, and reads the shim log before declaring anything a defect. ## Core Features & Use Cases - Guided installation: Detects existing installs, runs the installer from a clone or the skeleton's raw URL, and supports multiple named endpoints (vLLM, Ollama) with a configurable default. - Live verification: Exercises the full chain (launcher → shim → server → tool call) via print mode, then inspects the shim log as the source of truth for what was sent. - Troubleshooting playbook: Covers unreachable endpoints, Ollama context truncation, 400 role errors, oversized prompts from MCP tool schemas, and output-token cap conflicts. - Use Case: A developer gets a new vLLM box serving a Qwen model and wants Claude Code to run against it without losing Anthropic access on plain claude — this Skill installs the scoped launcher and confirms a tool-call round trip. ## Quick Start Ask the assistant to set up claude-local against your vLLM server at its base URL and verify it with a live print-mode run.

Frequently Asked Questions about setup-claude-local

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run Claude Code on a self-hosted vLLM model?▼

Install the claude-local launcher with install.ps1 -Upstream pointing at your server's base URL. It configures a local shim on 127.0.0.1:8098 that folds system-role turns into user turns and clamps max_tokens, then launches Claude Code scoped to that endpoint.

How do I verify claude-local is working after installation?▼

Run claude-local --list to confirm the endpoint is up, then run a print-mode command with a Bash echo tool call and check the JSON output for is_error false. Finally read the shim log tail to confirm a 200 response with expected prompt size.

Can I use claude-local with Ollama instead of vLLM?▼

Yes, add an Ollama endpoint with install.ps1 -Name ollama -Upstream http://127.0.0.1:11434, pinning -Model and setting -Context to match OLLAMA_CONTEXT_LENGTH. Ollama never reports its context length and silently truncates prompts that exceed it.

Why does Claude Code return a 400 role error against vLLM?▼

The 400 error means Claude Code is talking to vLLM without the shim, which folds role: system turns into user turns that vLLM accepts. Check that ANTHROPIC_BASE_URL is http://127.0.0.1:8098 in the session rather than pointing directly at the server.

Why does claude-local say the prompt is too long before I type anything?▼

MCP tool schemas dominate the first request — 11 servers can add roughly 113k tokens. Relaunch with claude-local --no-mcp or a minimal --mcp-config, trim oversized CLAUDE.md files, or raise --max-model-len on the server.

Does setting up claude-local affect the regular claude command?▼

No, the launcher scopes ANTHROPIC_BASE_URL to its own process, so plain claude continues using Anthropic. You should never export ANTHROPIC_BASE_URL globally, as that would redirect every session to the local server.

Related Skills

Explore More Software Engineering Skills→
inho-teaminho-team
6

Qlocal-llm-setup

Configure local LLM servers and integrate them with Claude Code.

Official
Advanced
ZardLi1115ZardLi1115
4

serving-llms-vllm

Deploy vLLM inference servers with OpenAI-compatible APIs and quantization.

Community
Intermediate
surfingaliensurfingalien
1

context-optimizer

Diagnose and reduce token usage through compaction, MCP audits, and subagent delegation.

Community
Intermediate
PlaidmustachePlaidmustache

searxng-search

Aggregate search results from multiple engines via a self-hosted SEARXNG instance.

Community
Intermediate