nerv-cli 0.1.0

Nervous Machine CLI — compose/deploy/sync a self-learning causal graph against the deployed NM MCP servers. Orchestration only: the learning math lives server-side and, for devices, in a compiled sha256-pinned kernel artifact. No kernel or framework source ships in this package.

nerv — the Nervous Machine CLI

Compose, deploy, and sync a self-learning causal graph against the deployed Nervous Machine servers. Domain experts write two YAML files; an agent orchestrates the rest; devices learn autonomously with a compiled kernel.

Two rules separate nerv from demo-ware:

1. Call the deployed servers or fail — never fabricate. No dry-run mode that prints plausible numbers. Misconfig/auth → it raises; a tool error → the real error surfaces.
2. The learning math lives server-side. nerv only orchestrates. Device deploys receive the kernel as a compiled, sha256-pinned artifact fetched with your own credential — never as source. The bundle's loader re-verifies the hash and replays a golden trajectory before first use.

Setup

pip install nerv-cli
nerv login          # OAuth sign-in with your issued User ID + access key

Operators with a static NM_API_TOKEN can set it in the environment instead; nerv login is the customer path (tokens cached, auto-refresh, pod-scoped server-side by your credential).

The lifecycle

nerv new my-domain --namespace <your-ns>   # scaffold the no-code kit: yamls + SKILL.md + AUTHORING.md
nerv compose prior my-domain --check-only  # conformance pass, then collaborative composition
nerv compose validation my-domain          # validation pipeline + REQUIRED scale test
nerv deploy my-domain --target pi          # gated bundle: compiled kernel, hash-pinned MANIFEST
nerv sync my-domain --device <bundle-dir>  # consume the device outbox

No learning before the calibration slice freezes the measurement scale — the deploy refuses without a passing scale test, on purpose. On the device, learning is autonomous (kernel only — no LLM, no network) and queues typed signals in results/outbox.jsonl: anomaly (acute) / curiosity (chronic) / promote (earned). sync is idempotent: promote writes back to your pod's global prior deterministically (dry-run by default; --apply-promotes writes); curiosity/anomaly become agent proposals for your review. Patches ship only new low-certainty edges — existing learned state is verified byte-identical or the deploy refuses.

Bring your own agent (MCP is the contract)

The deployed MCP servers are the fixed contract; the agent is swappable:

• Claude Code (recommended interactive): --emit claude-code prints a ready .mcp.json + prompt + run command. Without NM_API_TOKEN set, the config is header-free and your client discovers OAuth on first use.
• Any MCP-capable agent (Claude Desktop, Cursor, custom): --emit prints the prompt + server contract to wire in.
• Headless / CI: --agent anthropic drives Claude's API directly with your ANTHROPIC_API_KEY.

introspect is the observational loop (read-only): per edge it reads live certainty/weight and judges validating / revised / stalled — "I don't know yet" over a forced cause. Acting on a hypothesis is the separate curiosity/patch step, additive only.

Access & license

The CLI is Apache-2.0 (see LICENSE/NOTICE). The servers, the compiled kernel artifacts, and hosted graph state are separate licensed services — using them requires credentials issued by Nervous Machine, and every write is fenced to your pod namespace server-side.

Contact: [e-mail]