AXIOM is an open-source AI engineering platform that brings models, datasets, training, evaluation, runtime and deployment into one focused workspace. It is local-first, self-hosted, and built for engineers who want to own their AI stack.
Install from terminal · Website · Wiki · Roadmap · Contributing
╔══════════════════════════════════════════════════════════════╗
║ AXIOM CONTROL CENTER READY ◉ ║
╠══════════════════════════════════════════════════════════════╣
║ SYSTEM HEALTH ████████████████████ OPERATIONAL ║
║ AI RUNTIME ████████████████████ LOCAL ENGINE LIVE ║
║ MODEL REGISTRY ████████████████████ SYNCED ║
║ INFERENCE ████████████████████ ACTIVE ║
╠══════════════════════════════════════════════════════════════╣
║ ✓ Local workspace · Your models · Your data ║
╚══════════════════════════════════════════════════════════════╝
Modern AI development means juggling a different tool for every stage. AXIOM replaces the entire stack with one coherent engineering workspace.
| Stage | What AXIOM does | |
|---|---|---|
| ◎ | Models | Inspect metadata, manage a local registry, search Hugging Face |
| ◇ | Datasets | Validate, clean, deduplicate and analyse training data |
| ⌁ | Training | Hardware-aware plans, LoRA/QLoRA configs, fit estimation |
| ◌ | Evaluation | Benchmarks, quality checks and model comparison |
| ▣ | Runtime | Runtime foundations, MCP tooling and resource planning |
| ⌬ | Diagnostics | MCP tooling, system health, request tracing |
One project. One configuration. One workflow.
# macOS and Linux
curl -fsSL https://raw.githubusercontent.com/NetCore-Technologies/AXIOM-AI/main/installers/install.sh | bash
# Verify
axiom versionirm https://raw.githubusercontent.com/NetCore-Technologies/AXIOM-AI/main/installers/install.ps1 | iex
axiom versionAXIOM is intentionally installed and used from the terminal. The website does not expose binary download links.
# Start a project
axiom init my-ai
axiom status
axiom doctor
# Models
axiom model list
axiom model add my-model local --format safetensors
axiom model inspect ./models/my-model
axiom model search "mistral 7b"
# Datasets
axiom dataset inspect ./data/train.jsonl
axiom dataset validate ./data/train.jsonl
axiom dataset clean ./data/train.jsonl
axiom dataset stats ./data/train.jsonl
# Project
axiom project info
axiom project validate
axiom config show
# Hugging Face
axiom hf login
axiom hf status
# Training (plan generation)
axiom train plan 7
# Runtime
axiom mcp serveaxiom/
├── cli/ ← Typer command surface
├── core/ ← Platform engine
├── models/ ← Registry, inspection, metadata
├── datasets/ ← Validation, cleaning, stats
├── training/ ← Planning, hardware detection, fit estimation
├── evaluation/ ← Benchmarks, quality, comparison
├── runtime/ ← Provider integrations and runtime foundations
├── mcp/ ← Stdio MCP server and tools
└── config/ ← Configuration management
Intentionally modular. Each subsystem evolves independently without coupling to the rest.
AXIOM currently exposes its local backend through the Python CLI and a stdio MCP server (axiom mcp serve). This repository does not contain an HTTP service, database, CORS/auth middleware, or hosted API deployment. The Control Center therefore reports disconnected states until a real API contract is added; it does not invent live model, dataset, or runtime data.
Core Platform
- AXIOM project initialization
- Modular AI engineering architecture
- First-boot administrator setup
- Local authentication
- Session timeout protection
- Password management
- Modern AXIOM control center
- Light / dark UI support
CLI. 21 commands shipping
-
axiom version·axiom init·axiom doctor·axiom info·axiom status -
axiom model list·axiom model add·axiom model inspect·axiom model search -
axiom dataset inspect·axiom dataset clean·axiom dataset validate·axiom dataset stats -
axiom project info·axiom project validate -
axiom config show·axiom config validate - Hugging Face authentication · SuperCompress · Integration registry · MCP support
Models
- Model registry · inspection · metadata · format awareness · parameter count · quantization
- Automated download manager · conversion pipeline · benchmark suite · compatibility checks · version management
Datasets
- JSONL inspection · cleaning · validation · duplicate detection · statistics · field detection · token estimation
- Versioning · diffing · deduplication engine · sampling · augmentation · quality scoring
Training
- Training planning · hardware-aware planning · LoRA planning · CPU/GPU detection · model-fit estimation
- Training execution · job management · experiment tracking · checkpoints · dashboards · multi-GPU · distributed · hyperparameter search
Evaluation
- Evaluation subsystem foundation
- Automated pipelines · benchmark runner · dataset-based eval · model comparison · regression testing · custom metrics · reports
Runtime
- Runtime foundation · MCP foundation · integration registry
- Production serving · streaming inference · request batching · scheduling · autoscaling · health checks · load testing · optimization
Observability
- Diagnostics foundation · system information · CLI health checks
- Live telemetry · request tracing · token tracking · GPU monitoring · training telemetry · metrics · log search · profiling · export
Security
- Administrator authentication · session timeout · password management · GPG-signed releases
- API auth · secrets management · RBAC · audit logging · security diagnostics · enterprise controls
Developer Experience
- Python package · Typer CLI · release packaging · GitHub Actions pipeline · Wiki · community docs · issue templates · PR templates
- Python API docs · plugin SDK · shell completion · diagnostics bundle
- Agent-type questionnaire
- Use-case selection
- Privacy preference
- Latency preference
- Target tokens/sec
- Hardware-aware model planning
- Quantization recommendation
- Memory-fit estimation
- Hugging Face model discovery
- Real device benchmark engine
- Automatic quantization/export
- Benchmark → tune → re-run loop
- Model search
- Model metadata lookup foundation
- One-click model import
- Local model cache management
- Compatibility scoring
- Artifact verification
- Policy/safety indicator inspection
- Config inspection
- Read-only audit workflow
- Expanded policy metadata analysis
- Model lineage reporting
AXIOM does not remove or bypass model safety controls. The policy feature is an audit and transparency tool.
AXIOM can analyze a large Hugging Face or local model and create a device-aware runtime configuration for a specific agent workload.
- Coding Agent
- Reasoning Agent
- Research Agent
- General Assistant
- Automation Agent
- Math Agent
- Writing Agent
- Multilingual Agent
The optimizer:
- analyzes the model
- inspects CPU/RAM/GPU/VRAM
- chooses a quantization target
- creates a minimal runtime bundle
- keeps inference-critical configuration/tokenizer files
- retains model weights
- removes non-runtime repository artifacts
- configures agent-specific context/temperature settings
- uses a 10 tok/s target by default
- requires a real benchmark before claiming 10 tok/s achieved
axiom optimize profiles
axiom optimize run --model Qwen/Qwen3-8B --profile 1 --target-tps 10AXIOM does not delete arbitrary model knowledge from weights. Removing learned capabilities safely requires a model-conversion, distillation, pruning, or retraining workflow rather than file deletion.
AXIOM beta.5 adds a questionnaire-driven model optimization workflow.
- Coding Agent
- Reasoning Agent
- Research Agent
- General Assistant
- Automation Agent
- Math Agent
- Writing Agent
- Multilingual Agent
- Select the agent workload.
- Select a Hugging Face or local model.
- Choose a throughput target, with 10 tok/s as the default.
- Inspect the target system's CPU, RAM, GPU, VRAM, and disk.
- Estimate the model's memory footprint.
- Select a quantization target.
- Build a runtime-focused bundle.
- Benchmark locally where a supported runtime is available.
AXIOM does not claim 10 tok/s until real hardware benchmarking verifies it.
✅ Completed
Core platform · First-boot admin · Authentication · Session controls · Model registry · Dataset inspection & cleaning · Training plan generation · Hardware detection · Model-fit estimation · Hugging Face integration · SuperCompress · Integration registry · MCP foundation · CLI expansion (21 commands) · Project validation · Config validation · Diagnostics · Wiki · Community docs · Release packaging · GPG-signed releases
Next. AI Engineering
- Training execution · job manager · queue · experiment tracking · checkpoints · resume/recovery · hyperparameter search
- Evaluation pipelines · benchmark runner · model comparison · regression testing · reports
Next. Data and Models
- Dataset versioning · diffing · deduplication · quality scoring · sampling · augmentation · lineage
- Model download manager · conversion · compatibility matrix · benchmarking · version management · quantization workflows
Next. Runtime and Observability
- Production serving · streaming inference · batching · scheduling · autoscaling · health monitoring
- Live telemetry · request tracing · GPU monitoring · log search · performance profiling · metrics export
🔭 Platform Expansion
- Plugin SDK · Python API · CLI shell completion · remote training · multi-GPU · distributed training/inference · agent orchestration · advanced MCP · deployment automation
🌐 Long-Term Vision
- Full AI experiment workspace · end-to-end lifecycle management · collaborative AI engineering · enterprise deployment · AXIOM plugin marketplace · advanced agent platform
| Own your models | Use what you choose, on infrastructure you control |
| Own your data | Your datasets stay yours |
| Reproducibility first | Training and evaluation reproducible from config |
| Local-first | Local hardware is a first-class environment |
| Modular by design | Integrates with existing ecosystems, no lock-in |
- Questionnaire-driven planning
- Hugging Face discovery
- Hardware-aware planning
- Quantization recommendation
- Real per-device benchmark engine
- Automatic quantization/export
- Optimize → benchmark → retune loop
- Per-device performance profiles
- Benchmark history
- Performance regression detection
- Model compatibility scoring
- Model lineage
- Training execution
- Training job manager
- Training queue
- Experiment tracking
- Checkpoint management
- Resume/recovery
- Hyperparameter search
- Multi-GPU orchestration
- Distributed training
- Evaluation pipelines
- Benchmark runner
- Model comparison
- Regression testing
- Custom metrics
- Evaluation reports
- Evaluation dashboard
- Production serving
- Streaming inference
- Request batching
- Request scheduling
- Runtime autoscaling
- Endpoint management
- Runtime load testing
- Inference optimization
- Token throughput telemetry
- Latency telemetry
- CPU/RAM/GPU monitoring
- Request tracing
- Performance profiling
- Benchmark dashboards
- Metrics export
- Agent Model Optimizer
- Agent questionnaire
- Model Policy Audit
- Interactive hardware profiler
- Live benchmark panel
- Training workspace
- Evaluation workspace
- Runtime control center
- Browser credential storage cleanup
- Model Policy Audit
- API authentication
- Secrets manager
- Role-based access control
- Audit logging
- Security health dashboard
AXIOM is young. Architecture and APIs move fast.
Open an issue before large feature contributions. Bug fixes, tests, docs and tooling improvements are always welcome.
→ CONTRIBUTING.md · Code of Conduct · Security Policy
MIT. LICENSE
Built by @manit6752025 and contributors · netcore-technologies.github.io/AXIOM-AI
AXIOM. Build AI. Own AI.