Supply Chain

What Network-AI installs, what it executes, what it writes to disk, and what it calls over the network — so operators can make informed trust decisions.

1. Runtime Dependencies

Network-AI has one production runtime dependency:

That's it. The core library, all 32 adapters, the MCP server, and the blackboard are implemented with Node.js built-ins only.

BYOC design principle: Adapters are "Bring Your Own Client". They declare interfaces and types but import no LLM SDK, no HTTP library, and no vector store at require-time. The operator passes a pre-constructed client in. This means installing network-ai does not install OpenAI, LangChain, AutoGen, or any other framework SDK.

2. Development Dependencies (not installed in production)

devDependencies are not included in the published npm package.

3. What Runs at Install Time

There are no install-time scripts. npm install network-ai only downloads and unpacks the package; no code runs.

The prepublishOnly script (npm run build && npm run test:all) runs only when the maintainer publishes a new version. It is not executed on the operator's machine.

4. What Writes to Disk at Runtime

All writes are confined to the configured data directory. Path traversal is blocked: all file operations resolve the final path and assert it is within the data root before I/O.

5. What Calls Over the Network

Core library: zero outbound network calls

The core library (index.ts, lib/, security.ts) makes no outbound network calls. All network I/O is performed by:

• The operator's BYOC client (passed in by the operator at runtime)
• lib/mcp-transport-sse.ts, which receives inbound HTTP/SSE

connections — it does not initiate outbound connections

• The Streamable HTTP MCP transport (lib/mcp-transport-http.ts), which also receives

inbound connections — no outbound calls

Adapters: outbound calls only to operator-configured endpoints

The 32 adapters (adapters/) may make outbound HTTP calls, but only to endpoints explicitly provided by the operator at runtime (BYOC — Bring Your Own Client). No adapter hard-codes a URL that is called automatically at import or on first use.

Python helper scripts

The helper scripts (scripts/blackboard.py, scripts/check_permission.py, scripts/validate_token.py, scripts/revoke_token.py) make no network calls. They read and write local files only.

Test suite only (not included in the published package)

5a. Shell Execution Surface

Network-AI's AgentRuntime (lib/agent-runtime.ts) provides a sandboxed shell executor as an explicit, opt-in product feature. It is not activated by importing the library or starting the MCP server — it requires the operator to construct an AgentRuntime instance with an explicit allowedCommands allowlist.

How it is protected:

The shellAccess and shellExec alerts shown on the Socket.dev package page are expected and intentional for this class of tool. They reflect the presence of the shell sandbox, not a security defect.

5b. Telemetry

Network-AI's telemetry system (lib/telemetry-provider.ts) uses a Bring Your Own Telemetry (BYOT) design. The default provider is NullTelemetryProvider, which discards all spans and metrics without making any network calls.

Zero telemetry is emitted by default. No data is sent anywhere unless the operator explicitly constructs and registers an ITelemetryProvider implementation that forwards spans to their own OTel collector.

6. npm Provenance & SLSA

Every release is published via GitHub Actions CI with:

- name: Publish to npm
  run: npm publish --provenance --access public
  env:
    NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}

The --provenance flag publishes a signed SLSA Build Level 2 attestation to the Sigstore transparency log. This attestation links the published package to the exact Git commit and CI workflow run that built it.

Verifying provenance:

npm audit signatures
# or
npm install network-ai
npx sigstore verify node_modules/network-ai

The provenance statement includes:

• Source repository: https://github.com/Jovancoding/Network-AI
• Builder: GitHub Actions (ubuntu-latest)
• Build workflow: .github/workflows/ci.yml
• Git ref: refs/tags/v<version>

7. Signed Git Tags

All release tags are pushed from CI and correspond to a signed commit on main. The tag format is v<semver> (e.g. v5.7.2).

8. CodeQL Scanning

Every push and pull request to main triggers CodeQL analysis (JavaScript/TypeScript). Results are published to the GitHub Security tab. Open alerts are tracked and resolved before each release.

9. Package Contents (files in package.json)

The published npm package includes:

dist/          compiled JavaScript + declaration files (CJS, CommonJS target ES2022)
dist/esm/      ESM build (module: Node16, for bundlers and native ESM consumers)
bin/           CLI entry points (TypeScript source)
types/         additional .d.ts declarations
scripts/       Python helper scripts
socket.json    MCP socket configuration
README.md
QUICKSTART.md
INTEGRATION_GUIDE.md
SKILL.md
LICENSE

Notably excluded from the published package:

• test-*.ts files
• data/ directory
• .github/
• node_modules/
• TypeScript source files (other than bin/)

10. Dependency Update Policy

• Dependabot is enabled for npm (weekly PRs for patch/minor updates).
• All dependency update PRs run the full test suite (3,603 tests) before merge.
• Major version bumps require manual review and a CHANGELOG entry.
• New runtime dependencies require explicit approval and an update to this file.