richfrem/local-llm-setup
plugins/cli-agents/skills/local-llm-setup/SKILL.md
Cross-platform setup wizard for the local Gemma 4 12B inference stack. Automates llama-server installation (binary download or Metal/CUDA/Vulkan/ROCm compile), model download, routing proxy daemon install (launchd/systemd/NSSM), and Mode A/B validation. Covers Day 1 bootstrap and Day 2+ reconfiguration.
$ install --global
skillsauthnpx skillsauth add richfrem/agent-plugins-skills local-llm-setupInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Jul 13, 2026, 6:10 AM257.0s10 files scanned
SKILL.md
- name:
- local-llm-setup
- plugin:
- cli-agents
- description:
- >
- allowed-tools:
- Bash, Read, Write
<example>
<commentary>User wants to set up local Gemma 4 for the first time on a Mac.</commentary>
User: Set up local LLM with Gemma 4 on my M1 Mac
Agent: Detects Metal GPU, compiles llama-server from source, downloads gemma-4-12b-UD-Q4_K_XL.gguf,
starts server, installs routing proxy via launchd, validates with a Mode B timing test (~2s).
</example>
<example>
<commentary>User wants to test Mode B task delegation speed vs Mode A proxy.</commentary>
User: Compare Mode B vs Mode A speed for local Gemma
Agent: Runs `time python3 scripts/run_agent.py /dev/null /dev/null /tmp/t.md "hello" --cli llama`
(~2s), then `time claude --model gemma-4-12b -p "hello"` (~30–60s cold), reports the delta.
</example>
Co-located Scripts (
Primary Use Case: Mode B Task Delegation
Mode B is the fast path. run_agent.py sends a lean prompt directly to llama-server — no proxy overhead, no 29K system prompt. Measured: ~2s wall clock for a typical bounded task.
# Start llama-server (required for cli=llama)
python3 scripts/run_server.py
curl http://localhost:8089/health # must return {"status":"ok"}
# Mode B task delegation — fast path (~2s)
time python3 scripts/run_agent.py agents/refactor-expert.md target.py output.md \
"List the top 3 issues." --cli llama
# Mode B with custom max tokens
python3 scripts/run_agent.py /dev/null /dev/null /tmp/out.md \
"Summarize this architecture decision." --cli llama --max-tokens 300
Available agent personas (pass as PERSONA_FILE):
| Persona | Role |
|---------|------|
| agents/refactor-expert.md | Code quality — SOLID/DRY smell taxonomy |
| agents/security-auditor.md | OWASP vulnerability audit |
| agents/architect-review.md | C4/SOLID structural review |
| agents/red-team-reviewer.md | Adversarial exploit analysis |
| agents/compliance-reviewer.md | Coding standards drift detection |
| agents/pr-reviewer.md | Diff review — ship/hold decision |
| agents/test-writer.md | Unit test generation |
| agents/debate-synthesizer.md | Multi-perspective synthesis |
| agents/output-validator.md | Output guardrail / hallucination check |
| agents/self-critic.md | Reflection loop — task-fit check |
| agents/performance-analyst.md | Bottleneck and scale analysis |
Mode A (Optional — Interactive Proxy)
Mode A routes Claude Code itself through Gemma via a proxy. It carries ~29K tokens of system prompt overhead per session, making the first turn 30–60s. Not recommended for task delegation — use Mode B instead.
python3 scripts/enable_global_routing.py # install launchd/systemd/NSSM daemon
python3 scripts/disable_global_routing.py # remove daemon
Co-located Scripts (scripts/)
| Script | Purpose |
|--------|---------|
| run_server.py | Start llama-server (authoritative params) |
| run_agent.py | Task router — Mode B, 6 backends |
| enable_global_routing.py | Install Mode A proxy daemon |
| disable_global_routing.py | Remove Mode A proxy daemon |
| routing_proxy.py | Mode A API compatibility proxy (port 4000) |
Related Skills
richfrem/issue-worktree-agent
testing
VerifiedTrustedCommunity
Skill for creating and managing isolated git worktrees (`.worktrees/issue-NNN`) for issue execution branches. USE ONLY when setting up or cleaning up isolated git worktrees for specific issue execution. DO NOT USE for managing local task files (use `task-agent`) or escalating tasks to issues (use `github-issue-backlog-agent`).
4SKILL.mdUpdated Jul 29, 2026
richfrem/issue-pr-lifecycle-agent
data-ai
VerifiedTrustedCommunity
Skill for orchestrating the end-to-end GitHub issue lifecycle flow: Issue -> Worktree -> Implementation -> PR Creation -> Resolution Closure. USE ONLY when running or dry-running full lifecycle orchestration for resolving an issue with a PR. DO NOT USE for isolated worktree management only (use `issue-worktree-agent`) or logging issues (use `github-issue-agent`).
4SKILL.mdUpdated Jul 29, 2026
richfrem/github-issue-prioritizer
tools
VerifiedTrustedCommunity
Automatically ranks GitHub issues (P0-P3) based on friction tier, frequency, and blockages, synchronizing priority labels and GitHub Projects v2 custom fields.
4SKILL.mdUpdated Jul 29, 2026
richfrem/github-issue-backlog-agent
testing
VerifiedTrustedCommunity
Bridge skill for escalating ephemeral local task scratchpad items (`tasks/*.md`) into durable, taxonomy-validated, evidence-rich GitHub Issues. USE ONLY when promoting a single-session local task into durable repository backlog. DO NOT USE for managing local kanban boards (use `task-agent` instead) or directly querying/commenting on issues (use `github-issue-agent` instead).
4SKILL.mdUpdated Jul 29, 2026