Train your own AI with only your data. Claude Code-like power running entirely on your machine.
Your code. Your data. Your AI. Eldric Client is a complete AI development environment that never sends your data to the cloud. Train custom models, build knowledge bases, and automate your entire workflow — all locally.
Generic AI models don't understand your codebase, your architecture decisions, or your team's conventions. With Eldric Client, you can create an AI that truly knows your project.
Feed your entire repository, documentation, and coding standards to train a model that understands your specific patterns, naming conventions, and architectural decisions.
Whether it's financial regulations, medical terminology, or proprietary algorithms — train models that speak your industry's language fluently.
Every training step happens on your machine. Your proprietary code, trade secrets, and customer data never touch external servers.
As your project evolves, retrain your model. Keep your AI assistant current with the latest changes, deprecations, and new features.
Collect your code, docs, and examples
Set LoRA parameters and base model
Run training with progress monitoring
Merge adapter and start using
$ eldric
# Eldric v3.0 — Let's train a custom model
You: /train create
Eldric: Starting training job configuration...
[Training] Configure your fine-tuning job:
Base model: llama3.1:8b
Method: QLoRA (4-bit quantized)
Dataset: ./my-codebase-examples.jsonl
[Dataset] Analyzing 2,847 training examples...
✓ Code completions: 1,245
✓ Bug fixes: 834
✓ Documentation: 768
You: Start training with rank 32 and 3 epochs
[Training] Job started: job_abc123
Progress: ████████████░░░░░░░░ 62%
Loss: 0.847 → 0.234
ETA: 47 minutes
Training complete! Model saved to: ~/.eldric/models/my-custom-llama
You: /model my-custom-llama
Switched to my-custom-llama (fine-tuned on your codebase)
You: How do we handle authentication in this project?
Eldric: Based on your codebase patterns, you use JWT with refresh
tokens. The auth flow is in src/auth/middleware.ts. Here's how
to add a new protected route following your conventions...Why choose between models when you can merge them? Combine a coding specialist with a documentation expert, or blend reasoning capabilities with domain expertise.
MODEL MERGING WORKFLOW
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Code Model │ │ Docs Model │ │ Domain Model │
│ (CodeLlama) │ │ (Mistral) │ │ (Your LoRA) │
└────────┬────────┘ └────────┬────────┘ └────────┬────────┘
│ │ │
└───────────────────────┼───────────────────────┘
│
┌─────────▼─────────┐
│ MERGE ENGINE │
│ │
│ • SLERP (smooth) │
│ • TIES (sparse) │
│ • DARE (random) │
│ • Linear blend │
└─────────┬─────────┘
│
┌─────────▼─────────┐
│ YOUR CUSTOM │
│ SUPER MODEL │
│ │
│ Best of all │
│ three models! │
└───────────────────┘
You: /merge create
Eldric: Let's create a merged model.
[Merge] Select models to merge:
1. codellama:13b (coding specialist)
2. mistral:7b (general reasoning)
3. my-domain-lora (your fine-tune)
[Merge] Choose merge strategy:
→ SLERP - Spherical interpolation (recommended)
[Merge] Set weights:
codellama: 0.4
mistral: 0.3
my-domain: 0.3
[Processing] Merging tensors...
Progress: ████████████████████ 100%
Merge complete!
Output: my-merged-expert
Size: 13.2 GB
# Your new model combines:
# - CodeLlama's coding ability
# - Mistral's reasoning
# - Your domain knowledgeNot everything needs fine-tuning. For rapidly changing information, documentation, or large knowledge bases, RAG provides instant context injection without retraining.
PDF, Markdown, code files, HTML, DOCX, TXT — 20+ formats automatically chunked and embedded.
Find relevant context by meaning, not just keywords. FAISS and Chroma backends for fast retrieval.
Update your knowledge base anytime. No retraining needed — just re-index changed documents.
You: /learn ./company-docs
Eldric: Indexing documents for RAG...
[Learn] Scanning directory...
Found 342 documents (127 MB)
[Embed] Creating embeddings...
Progress: ████████████████████ 100%
Chunks: 8,456
Vector DB: FAISS
Knowledge base ready!
You: What's our policy on API rate limiting?
Eldric: [RAG Search] Found 3 relevant documents
Based on your internal documentation (api-guidelines.md):
Your API rate limiting policy is:
- Free tier: 100 requests/minute
- Pro tier: 1000 requests/minute
- Enterprise: Custom limits
Implementation uses token bucket algorithm with Redis backend.
See src/middleware/rate-limiter.ts for the implementation.
[Source: docs/api-guidelines.md, lines 45-78]Works completely offline. Perfect for classified environments, secure facilities, or anywhere without internet.
Your training data, prompts, and outputs never leave your machine. Full compliance with data residency requirements.
Complete logging of all AI interactions. Know exactly what your AI assistant did and why.
New to a project? Train a model on the codebase, then ask it anything: "How does authentication work?", "Where are the API routes?", "What's the deployment process?"
Fine-tune on your existing docs and coding style. Generate consistent, accurate documentation that matches your team's conventions.
Train on your code review history. Get suggestions that align with your team's standards and catch issues specific to your codebase.
Feed your legacy codebase to create an AI that understands that old Perl script or COBOL system nobody wants to touch.
See how Eldric Client can transform your development workflow.
Request Demo