Stop renting AI. Own it.
Your team already has thousands of AI conversations. Brewmode turns them into a local model that handles 80-90% of routine tasks for free. Complex work still goes to frontier APIs.
$ brewmode distill
[1/4] Ingesting sources...
Claude Code: 847 conversations
Cursor: 234 conversations
ChatGPT: 612 conversations
[2/4] Anonymizing...
Stripped 4,187 entities (0.94 avg confidence)
[3/4] Curating...
Selected 1,299 / 1,693 above threshold 0.5
[4/4] Training Qwen3-8B + LoRA...
Step 600/600 | Loss: 0.42 | ETA: done
Model saved: brewmode-v1.Q4_K_M.gguf (4.7 GB)
Four steps. Zero ML expertise.
Every step happens locally. Your data never leaves your machine.
Ingest
Drag-and-drop your conversation exports or point to local data directories. Supports 7 AI tools out of the box.
Anonymize
5-layer PII stripping: file paths, API keys, internal URLs, company names, and code-specific secrets. Powered by Presidio.
Curate
Score every conversation 0.0-1.0 based on code quality, completeness, corrections, and tool usage. Filter and deduplicate.
Train
Fine-tune with LoRA on your curated dataset. Export as a quantized GGUF model ready to run with Ollama or llama.cpp.
Smart routing. Massive savings.
Most AI tasks are routine. Route them to your local model for free. Send only the hard problems to frontier APIs.
80-90%
Local model
Free
10-20%
Frontier API
Paid
Before Brewmode
100% of requests go to frontier models at $0.015-0.06/1K tokens
After Brewmode
80-90% of requests handled locally for free. Only complex tasks hit paid APIs.
Actual savings depend on your usage patterns and task complexity distribution.
Your conversations are already the training data.
Auto-detect local installations or upload exported archives. Seven sources, one unified schema.
Claude Code
auto-detect~/.claude/
JSONL logs
Claude.ai
Settings export
Export ZIP
ChatGPT
Data Controls
Export ZIP
Gemini
Google Takeout
Takeout ZIP
Cursor
auto-detect~/.cursor/
SQLite + JSON
Codex CLI
auto-detect~/.codex/
Session logs
Cline
auto-detect~/.vscode/
VS Code data
Unified Schema
interface ConversationTurn {
id: string;
source: 'claude_code' | 'claude_chat' | 'chatgpt' | 'gemini' | 'cursor' | 'codex' | 'cline';
role: 'user' | 'assistant' | 'tool_call' | 'tool_result';
content: string;
metadata: { has_code: boolean; token_count: number; language?: string };
}5-layer anonymization. Built for developers.
Goes far beyond standard PII detection. Custom recognizers for API keys, infrastructure secrets, and code-specific patterns.
Standard PII
Names, emails, phone numbers, addresses via Microsoft Presidio
API Keys & Tokens
sk-*, pk_*, AKIA*, ghp_*, glpat-*, Bearer tokens
Infrastructure
Internal URLs, DB connection strings, AWS ARNs, private IPs
Code Paths
File paths with usernames, internal package names, git remotes
Business Context
Company names, product references, proprietary terminology
Fix the API call to https://api.wdc.internal/v2/pricing using key sk-proj-aK7x...mN9q Deploy to staging: ssh venkat@10.0.1.42 "cd /opt/dema/api && git pull" File: /Users/venkat/wd-tools/src/api.ts
Fix the API call to [INTERNAL_URL_1]/v2/pricing using key [API_KEY_1] Deploy to staging: ssh [USER]@[PRIVATE_IP_1] "cd [DEPLOY_PATH_1] && git pull" File: /Users/[USER]/[PROJECT_1]/src/api.ts
See the results.
Side-by-side comparisons of frontier, base, and Brewmode-trained models on real coding tasks.
Visual Comparison
30 coding prompts evaluated across frontier, base, and fine-tuned models. See code output quality side by side.
View evaluationAmazon-Style Comparison
Product-style comparison of model capabilities, cost, and quality metrics across categories.
View comparisonAI Router Demo
Live routing demo: simple tasks go to your local model (free), complex tasks go to frontier (paid).
Try the routerGet started in 5 minutes.
Drag and drop your conversation exports. One click to train. Auto-deploy to Ollama.
Start Building