Local Ollama LLM Offloading

You have access to a local Ollama instance at http://localhost:11434/ for offloading tasks that don't need Claude's full capabilities. This saves token credits for complex reasoning, multi-step planning, and tool-heavy workflows.

When to Use Ollama

ALWAYS prefer Ollama for these tasks (saves ~90% of credits for routine work):

• Summarizing individual files (purpose, exports, structure)
• Reviewing code for bugs, style issues, or security problems
• Analyzing git diffs before committing
• Quick syntax/API questions ("What does X do in TypeScript?")
• Generating boilerplate (simple CRUD, test stubs, type definitions)
• Explaining grep/search results in context
• Pre-screening large codebases before deep Claude analysis
• Drafting commit messages, PR descriptions, or documentation
• Generating Zod schemas, Drizzle table definitions from descriptions
• Writing test cases from function signatures

Use Claude (not Ollama) for:

• Multi-file architectural decisions requiring conversation context
• Complex debugging requiring tool access (running tests, reading errors)
• Tasks requiring web search, browser interaction, or external APIs
• Security-sensitive financial logic (payment calculations, Stripe integration)
• Anything requiring previous conversation context

How to Use

The helper script is at .claude/skills/ollama/ollama.sh. Run via Bash:

# Check availability and which model is loaded
.claude/skills/ollama/ollama.sh status

# Quick question (fast, no reasoning ~0.5-2s warm)
.claude/skills/ollama/ollama.sh query "What does pgEnum do in Drizzle?"

# Deep question (with reasoning, slower ~10-30s)
.claude/skills/ollama/ollama.sh think "Should I use a db transaction for this?"

# Summarize a file
.claude/skills/ollama/ollama.sh summarize server/services/payout.ts

# Review code for issues
.claude/skills/ollama/ollama.sh review components/shared/captcha.tsx

# Analyze a diff
git diff --staged | .claude/skills/ollama/ollama.sh diff

# Explain grep matches
.claude/skills/ollama/ollama.sh grep-explain "calculateApplicationFee" server/services/payout.ts

Graceful Degradation

• Server may not be running. If Ollama is unavailable, fall back to Claude silently — never error out to the user or ask them about it. Just do the work yourself.
• The model changes over time. The script auto-discovers whatever model is available. Don't hardcode model names. Current models use a thinking mode that the script handles automatically.
• Cold start: ~5-10s. First request after idle loads the model to GPU. Run status early in a session to warm up. Subsequent requests are fast (~0.5-2s).
• Not a replacement for Claude. Ollama handles routine read/analyze tasks but lacks tool access, conversation context, and Claude's reasoning depth. Use it as a first pass.

Batch Workflow Patterns

Explore a directory (instead of reading every file yourself)

for f in server/services/*.ts; do
  echo "=== $f ==="; .claude/skills/ollama/ollama.sh summarize "$f"; echo
done

Pre-review before committing

git diff --staged | .claude/skills/ollama/ollama.sh diff

Generate test stubs

.claude/skills/ollama/ollama.sh query "Write bun:test stubs for: export function getEffectiveShareRate(clinic: { revenueShareBps: number; foundingClinic: boolean; foundingExpiresAt: Date | null }): number"