Personal compute adapter for SpecGraph and SpecPM.
SpecNode lets users and open-source maintainers run repository analysis and LLM-backed specification extraction on their own devices while SpecGraph and SpecPM provide orchestration, validation, registry, provenance, and observability.
License: MIT. See LICENSE.
This repository is at the specification bootstrap stage. There is no installable runtime yet.
The first implementation target is a local CLI or daemon that connects to a SpecGraph/SpecPM control plane through an outbound job channel and executes a small allowlisted set of typed jobs.
DocC documentation is published through GitHub Pages at:
https://0al-spec.github.io/SpecNode/
Build the static documentation locally:
make docs-buildSpecGraph / SpecPM control plane
|
| outbound WSS session from node
v
specnode on user's device
|
| localhost only
v
Ollama / LM Studio / llama.cpp / custom OpenAI-compatible endpoint
SpecNode should return structured artifacts:
specgraph.jsonspec-package.yamlprovenance.jsonusage_receipt.json
It should not return unrestricted local filesystem state, raw secrets, or unbounded LLM transcripts.
The smallest useful slice is:
- pair a local node with a user account
- discover one local OpenAI-compatible provider
- execute one SpecGraph extraction job
- execute one SpecPM package-build job
- validate structured output against schemas
- produce a usage receipt with provider, model, duration, token counts, and artifact hash
Target CLI shape:
specnode connect --code 8K4P-X2Q9
specnode providers list
specnode test --model qwen3:4b
specnode specpm build --repo . --model qwen3:4b
specnode specpm validate ./dist/spec-package.yamlspecpm.yaml SpecPM package manifest for SpecNode itself.
docs/
proposals/ Product and architecture proposals.
specs/ Protocol, security, and artifact contracts.
src/ Runtime implementation, once selected.
tests/ Conformance and regression tests, once runtime exists.
tools/ Developer or validation tooling.
- Outbound-only connection from node to service.
- Typed jobs only.
- No arbitrary shell from remote jobs.
- No LLM tool execution in the MVP.
- Repository content is untrusted input.
- Deterministic analyzers should run before LLM inference where practical.
- Artifacts should be hashable, schema-valid, and provenance-linked.
- SpecGraph: executable product ontology and governed intent graph.
- SpecPM: intent-level package manager and registry for specification bundles.
- Agent Passport: future identity and policy envelope for verifiable local agents.