Axon

A local neural switchboard for LLM traffic.

Axon discovers provider credentials already living on your machine, validates them without exposing raw secrets, and opens one crisp localhost gateway for OpenAI- and Anthropic-compatible clients.

Signal deck / Quickstart / Control surface / Security model

scan -> fingerprint -> vault -> route

---

Signal Deck

Modern LLM work is no longer "pick one model and pray." It is a live routing problem: premium models for deep work, fast models for search, cheap models for bulk passes, and provider keys scattered across tools that all expect different shapes.

Axon turns that sprawl into one local control surface.

Old workflow	Axon workflow
Paste keys into every client	Reuse credentials already configured on the machine
Guess which provider a key belongs to	Resolve by base URL, distinctive prefix, then env name
Maintain separate proxy setups	Point OpenAI and Anthropic SDKs at one localhost gateway
Wonder whether a key authenticates	Probe the provider's own endpoint with zero-cost validation
Risk secrets in logs and screenshots	Display fingerprints only

The name comes from the axon: the fiber that carries a neuron's output signal to the right downstream target. Axon does the same job for model traffic.

Quickstart

pip install -e .
axon scan
axon scan --validate

Run the gateway:

pip install -e ".[server]"
axon serve

Point clients at Axon:

# OpenAI-compatible clients
base_url=http://127.0.0.1:4000/v1
model=gpt-4o

# Anthropic-compatible clients, including Claude Code
base_url=http://127.0.0.1:4000
model=claude-sonnet-4-6

Control Surface

Command	What it does	Secret behavior
axon scan	Finds configured providers from env vars and known tool configs	Prints fingerprints only
axon scan --validate	Checks which keys authenticate	Calls only the resolved provider endpoint
axon serve	Starts the dual-ingress gateway on 127.0.0.1:4000	Keeps usable keys in memory only
AXON_API_KEY=... axon serve --host 0.0.0.0	Enables authenticated non-localhost serving	Requires inbound Authorization or x-api-key

Served endpoints:

Surface	Endpoint	Streaming
OpenAI models	GET /v1/models	n/a
OpenAI chat	POST /v1/chat/completions	Yes
Anthropic messages	POST /v1/messages	Yes
Health	GET /healthz	n/a

Routing Map

%%{init: {"theme": "base", "themeVariables": {"background": "#0A0F12", "primaryColor": "#11161A", "primaryTextColor": "#F3F6E8", "primaryBorderColor": "#74D3AE", "lineColor": "#E6C86E", "secondaryColor": "#191713", "tertiaryColor": "#171A22"}}}%%
flowchart LR
    client["SDKs / CLIs / editors"] --> ingress["Axon localhost gateway"]
    env["Env vars"] --> discovery["Discovery engine"]
    files[".env + tool configs"] --> discovery
    cursor["Cursor state.vscdb"] --> discovery
    discovery --> vault["In-memory secret vault"]
    vault --> ingress
    ingress --> egress["LiteLLM egress"]
    egress --> openai["OpenAI"]
    egress --> anthropic["Anthropic"]

    classDef local fill:#0E1516,stroke:#74D3AE,color:#F3F6E8;
    classDef secure fill:#18150E,stroke:#E6C86E,color:#F3F6E8;
    classDef provider fill:#171A22,stroke:#8EA7FF,color:#F3F6E8;
    class client,ingress,discovery local;
    class vault secure;
    class egress,openai,anthropic provider;

Loading

Roadmap

Milestone	Status	Scope
M0 Discovery engine	Done	Env vars, Windows registry, .env, known tool configs, Cursor SQLite, provider detection, fingerprint-only output
M1 Dual ingress gateway	Done	OpenAI-compatible and Anthropic-compatible APIs backed by LiteLLM, localhost by default
M2 Router	Next	Static role-based routing, then cheap-first cascades
M3 Dashboard	Next	Active Models view, discovery cards, and cost stats

Discovery recognizes OpenRouter, Anthropic, DeepSeek, Google Gemini, Mistral, Groq, xAI, Together AI, Cohere, Perplexity, Azure OpenAI, and OpenAI. The M1 serving path currently loads OpenAI and Anthropic as first-class routed providers.

Security Model

Axon is a key-aware tool, so the security model is not a footnote. It is the shape of the product.

Rule	Behavior
Read-only discovery	Config files and databases are opened only for inspection
No raw-key output	CLI, logs, and display paths use fingerprints only
Provider-native probes	Validation calls only the resolved provider endpoint
In-memory serving vault	axon serve holds usable keys only inside the running process
Localhost by default	Non-localhost binds are refused unless AXON_API_KEY is set
Egress hardening	Client-supplied api_key, api_base, base_url, and related steering fields are stripped before provider calls

Read the full model in SECURITY.md.

Development

pip install -e ".[dev]"
python -m pytest -q

License

MIT.