Know your model, your data, and your machine before you spend a single GPU hour.

Quick start · Install · Commands · Website · Contributing

_{Sample output. Plans and estimates are not benchmark results. See the Web UI below.}

What is AXIOM?

AXIOM is a local-first CLI for answering the questions that come before a local AI job:

• What is in this model directory?
• Is this JSONL dataset structurally usable?
• What hardware is available on this machine?
• What is a reasonable starting plan?

It inspects local files, keeps small amounts of project metadata, reports what is available on the current machine, and suggests the next command. The terminal is the product. The optional daemon exposes the same local context as small JSON responses on a loopback address.

Project configuration lives in axiom.yaml; the local model registry lives in .axiom. Core inspection and planning do not require an AXIOM account.

AXIOM does not execute training, run model inference, or provide a hosted workspace. Estimates and plans are labeled as estimates and plans.

What is new in v0.3.0-beta.1

• Version bumped to v0.3.0-beta.1 across the package, badges, and docs.
• Refreshed README with an animated banner and a terminal walkthrough.
• Step-by-step quick start and a first-10-minutes guide below.

🚀 Quick start

Do these in order. Each step is one command.

1. Install AXIOM (see Install), then confirm it works:

   axiom version

2. Get oriented. None of these start the daemon:

   axiom --help
   axiom guide
   axiom summary

3. Create a project and check it:

   axiom init my-ai
   cd my-ai
   axiom project validate
   axiom config validate

4. Look at your machine:

   axiom system info

5. Get a starting plan for a 7B model:

   axiom train plan 7 --method qlora

Want the interactive mode? Run bare axiom. In an interactive terminal it prints the startup mark once per session, shows the useful command path, starts the local daemon on a free port, and stays attached until Ctrl-C. To open the graphical Control Center with the workspace API, run axiom gui; axiom ui is an alias. Both services bind to loopback and stop with Ctrl-C. The local API contract and compatibility policy are documented at docs/API.md; the live OpenAPI document is served at /api/openapi.json while the daemon runs. Saved training-plan snapshots are available in the GUI history and with axiom experiment list or axiom experiment show <id>; they are plans, not executed runs. To save an exact CLI plan snapshot, add --save; --json emits the plan and saved record for scripts:

axiom train plan 7 --method qlora --save --name "7B baseline" --tag baseline --json

Saved GUI workflow designs can be inspected and structurally validated from the same workspace store with axiom workflow list, axiom workflow show <id>, and axiom workflow validate <id>. These commands do not execute workflow steps. axiom evaluation run ./data/eval.jsonl --prediction-field prediction --reference-field reference scores precomputed outputs locally without model inference; see the evaluation guide for metric definitions.

🧭 First 10 minutes

Pick the path that matches what you have in front of you.

I have...	Run this	You get
A local model folder	axiom model inspect ./models/my-model	Format, files, architecture hints, weight size, conservative fit info.
A model name on Hugging Face	axiom model info org/model	Published metadata without downloading weights. Needs network.
A JSONL dataset	axiom dataset validate ./data/train.jsonl	Invalid records and duplicates; exits non-zero if records are invalid.
A dataset I want sized up	axiom dataset stats ./data/train.jsonl	Record, field, and rough token counts.
A messy dataset	axiom dataset clean ./data/train.jsonl ./data/train.cleaned.jsonl	A new cleaned file; the source is left unchanged.
A new machine	axiom system info	OS, architecture, CPU, RAM, GPU, VRAM, CUDA availability.
A plan to make	axiom train plan 7 --method qlora	Hardware-aware starting plan. It does not start training.
Some dev tools installed	axiom tools doctor	Executable and credential-reference presence, not vendor login state.

📦 Install

Debian / Ubuntu / Linux Mint (APT, amd64)

AXIOM publishes a signed APT repository alongside the project website. Add its public key and repository, then install the axiom-ai package:

curl -fsSLo /tmp/axiom-archive-keyring.gpg https://netcore-technologies.github.io/AXIOM-AI/apt/axiom-archive-keyring.gpg
sudo install -Dm0644 /tmp/axiom-archive-keyring.gpg /usr/share/keyrings/axiom-archive-keyring.gpg
echo 'deb [arch=amd64 signed-by=/usr/share/keyrings/axiom-archive-keyring.gpg] https://netcore-technologies.github.io/AXIOM-AI/apt stable main' | sudo tee /etc/apt/sources.list.d/axiom.list
sudo apt update
sudo apt install axiom-ai

The APT package currently targets 64-bit Intel/AMD (amd64). The repository metadata and release signature are verified by APT using the dedicated public keyring.

PyPI

The beta package installs the axiom command from PyPI:

python -m pip install --upgrade --pre axiom-all
axiom version

axiom-all is the distribution name; the installed command is axiom. Stable releases can be installed without --pre once one is published.

macOS and Linux

curl -fsSL https://raw.githubusercontent.com/NetCore-Technologies/AXIOM-AI/main/installers/install.sh | bash
axiom version

The installer puts the command in ~/.local/bin/axiom.

• Linux x86_64 uses the latest published linux-x64 tarball and verifies its entry in SHA256SUMS before activation. The unpacked bundle is the faster repeated-start path.
• macOS and non-x86 Linux use an isolated Python environment and install from the repository's main branch. This fallback requires Python 3.11+ and is a source install, not a pinned native binary.

If ~/.local/bin is not on PATH, the installer prints the export command. Open a new shell after adding it.

Windows PowerShell

irm https://raw.githubusercontent.com/NetCore-Technologies/AXIOM-AI/main/installers/install.ps1 | iex

The installer downloads the latest non-draft Windows x64 release to %LOCALAPPDATA%\AXIOM, adds that directory to the user PATH, and asks you to restart PowerShell. Verify with:

axiom version

Check the Releases page for the exact assets and checksums available for a release.

🔌 The daemon boundary

Run the daemon by itself when a script or local integration needs HTTP:

axiom daemon --host 127.0.0.1 --port 0

Port 0 asks the operating system for an available port. AXIOM prints the resulting loopback URL, such as http://127.0.0.1:53142. While that process is running, the daemon provides read-only JSON routes:

GET /health
GET /api/info
GET /api/summary
GET /api/actions
GET /api/hardware
GET /api/tools

The default bind is local-only. A non-loopback bind is rejected unless --allow-network is supplied deliberately. The daemon is not an account service, hosted AXIOM backend, or general inference server. Stop it with Ctrl-C.

🖥️ Web UI

The release packages and the PyPI wheel bundle the web UI. Run axiom gui to start the workspace API and open the Control Center in your browser. Use axiom gui --no-open to print the local URL without opening a browser. The separate axiom daemon command remains available for the read-only JSON API.

_{Sample data. Sizes and plans are estimates, not benchmark results.}

• Quantization Lab compares FP16, INT8 and INT4 size estimates against the memory AXIOM detects, so you can see what fits before you download anything.
• Optimizer turns that into a conservative starting plan.

🛠️ Commands by job

Project and environment

axiom version
axiom guide
axiom summary
axiom summary --json
axiom info
axiom status
axiom doctor

guide explains the current directory and suggests safe next commands. summary returns a read-only snapshot of the project, hardware, visible tools, and one recommendation; --json is intended for scripts. doctor checks common executables and AXIOM package metadata.

Models

axiom model list
axiom model add my-model local safetensors
axiom model inspect ./models/my-model
axiom model search llama --path models
axiom model info org/model
axiom model analyze org/model
axiom model pull org/model

The first three local commands do not need network access. Hugging Face metadata and pull commands do; private or gated repositories also need axiom hf login. model pull performs disk checks before downloading.

Datasets

axiom dataset inspect ./data/train.jsonl
axiom dataset validate ./data/train.jsonl
axiom dataset stats ./data/train.jsonl
axiom dataset clean ./data/train.jsonl ./data/train.cleaned.jsonl
axiom dataset profile ./data/train.jsonl --duplicate-key id --duplicate-rate-threshold 0.05 --text-fields prompt,completion
axiom dataset profile ./data/train.jsonl --scan-pii
axiom dataset manifest ./data/train.jsonl
axiom dataset manifest ./data/derived.jsonl --parent-manifest-sha256 <source-sha256>
axiom dataset preview-clean ./data/train.jsonl

The current dataset path supports .jsonl. Validation checks JSON objects and duplicates; it does not establish semantic quality or training readiness. Bounded profiling reports UTF-8/newline/malformed-record diagnostics, configurable duplicate-key rates, aggregate whitespace-token length bins, and sensitive-field name warnings. Optional --sensitive-rule values add substring rules, while --ignore-sensitive-field excludes exact names to manage false positives. Reports contain structural counts and field names only; they never contain row values. PII pattern checks are off unless --scan-pii is supplied (or the GUI opt-in is checked); their email, phone-like, and Luhn-valid card matches are heuristics, not a guarantee. dataset manifest hashes at most the configured limit and labels prefix hashes partial; --parent-manifest-sha256 links a derived dataset to its source manifest. dataset preview-clean reports removal/normalization counts without writing any file; dataset clean remains a separate explicit command.

Hardware and planning

axiom system info
axiom train plan 7 --method auto
axiom train plan 7 --method lora
axiom train plan 7 --method qlora
axiom train plan 7 --method full
axiom ai plan --model ./models/my-model

The planner uses detected hardware and conservative assumptions. Its fit and VRAM values are planning estimates, not measured speed, memory usage, or a promise that a later runtime will succeed.

Experiment plans

axiom experiment list --query baseline --tag review --json
axiom experiment show <experiment-id>
axiom experiment update <experiment-id> --name "Reviewed plan" --tag baseline --tag approved
axiom experiment delete <experiment-id> --yes

Experiment records are immutable plan snapshots with separately editable names, notes, and tags. Search, exact-tag filtering, and pagination are available in the Control Center and CLI. Deleting a record does not delete model or dataset files; AXIOM does not yet create run artifacts.

Developer tools

axiom tools list
axiom tools list --json
axiom tools doctor
axiom tools plan opencode
axiom tools install opencode       # preview only
axiom tools install opencode --yes # explicit package-manager install

The catalog covers Codex, Claude Code, Antigravity CLI, GitHub Copilot CLI, Freebuff, Cursor Agent, free-pi, OpenCode, Gemini, OpenRouter, z.ai GLM, and a keep-awake helper. tools list and tools doctor observe the current machine; they do not prove that a vendor account is authenticated. Installation is a preview unless --yes is supplied. AXIOM does not silently execute remote installer scripts or store API-key values.

Headroom is a separate, optional local context-compression proxy. If it is installed independently, its own command path is:

headroom doctor
headroom proxy
headroom dashboard

AXIOM's role is limited to discovering or explaining Headroom. It does not bundle, install, start, or supervise it; only the user starts the proxy or dashboard. Do not read the integration as a promise of savings or a measured result.

MCP

Start the local stdio server from an MCP client configuration:

axiom mcp serve

It exposes local model, dataset, hardware, training-plan, and optional SuperCompress tools as a process. It is separate from the HTTP daemon. MCP tools are denied by default. Inspect the active policy, explicitly grant only the tools you need, or create and select a named profile:

axiom mcp permissions
axiom mcp allow axiom_system_info --profile default
axiom mcp profile create read-only
axiom mcp allow axiom_model_list --profile read-only
axiom mcp profile use read-only

The Control Center also edits those same local profiles. Dataset profiling in the GUI offers quick, standard, and deep bounded scan budgets; profile results include the effective limits. The dataset quality panel compares schemas across two or three files, checks required fields and expected JSON types, and downloads a values-free JSON report. It also exposes encoding/newline diagnostics, duplicate-key thresholds, token-length bins, and configurable privacy-name rules. PII scanning is an explicit opt-in and returns only category/field counts. Dataset manifests support full/partial checksums and parent-hash lineage links; cleaning previews never write files; the actual cleaner remains a separate CLI operation. Dataset reports exist only in the current GUI session and the page's Clear reports action removes them. Its Workflows page validates and saves versioned local graph designs, includes templates and revision history, and supports duplication. Saved workflows can be dry-run for resource estimates and an exportable software/seed manifest; all steps remain marked not-run and no job is created because workflow execution is not available. Existing saved designs autosave after edits by default (with a GUI toggle); new designs still require an explicit save.

The Models page supports local-folder discovery under models/, bounded inspection summaries, stable registry IDs, tags, favorites, pagination, and versioned JSON import/export. Removing an entry removes only registry metadata; it never deletes model weights.

✅ What is shipped, and what is not

Shipped in the beta

• A Python package and Typer CLI with a one-command local entry point.
• Project scaffolding, configuration checks, and local status/summary output.
• Model registration, local inspection, Hugging Face metadata analysis, and disk-checked model pulls.
• JSONL inspection, validation, duplicate detection, statistics, and cleaning.
• Hardware detection and hardware-aware LoRA, QLoRA, and full-training plans.
• A read-only loopback daemon with health, summary, action, hardware, and tool routes.
• MCP tools and an opt-in developer-tool catalog with preview-first installs.
• A static landing page that explains the terminal workflow.

Not shipped by this product boundary

• Training execution, job management, checkpoints, or run/outcome tracking. The GUI stores planning snapshots only; it does not claim they were run.
• Model inference, production serving, streaming, batching, or an OpenAI-compatible server.
• Live telemetry, background monitoring, hosted workspaces, or accounts.

🗂️ Repository shape

axiom/
├── cli/          Typer command surface
├── core/         Project, hardware, storage, and integration helpers
├── models/       Local registry, inspection, and metadata analysis
├── datasets/     JSONL inspection and cleaning
├── training/     Hardware-aware plan generation
├── optimizer/    Agent workload plans and runtime bundles
├── runtime/      Optional runtime integrations
├── daemon.py     Dependency-free loopback HTTP daemon
└── mcp/          Stdio MCP server and tools

The public website is intentionally static. Product work belongs in the CLI, local daemon, and reproducible project files.

🧱 Principles

• Local first: keep models, datasets, configuration, and first-pass decisions close to the machine that will run them.
• Evidence before execution: expose assumptions before a user spends time or compute.
• Explicit side effects: preview external tool setup and require opt-in for package installation.
• Honest boundaries: separate detections, estimates, plans, and measured results.

🤝 Contributing

AXIOM is in active beta development. Bug fixes, tests, documentation, and tooling improvements are welcome.

Read CONTRIBUTING.md before making a larger change.

📄 License

MIT. See LICENSE.

Built by Manit Arora and Rangan V Balaji.

Website