AI Power Grid Provider Integrations

This workspace turns AIPG's public /v1 API into tested integrations for frameworks developers already use. It starts with one reusable conformance runner so every provider contribution proves the same baseline instead of shipping a different hand-written smoke test.

Integrations

• dify-aipg/ - native Dify LLM provider for the Grid's live OpenAI-compatible text models.
• elizaos-aipg/ - ElizaOS text model provider plus explicit image, video, and music actions.
• open-webui-aipg/ - Open WebUI direct/admin connection guide and upstream documentation draft.
• ai-sdk-aipg/ - native Vercel AI SDK v7 provider for text, image, experimental video, and typed music generation.
• langchain-aipg/ - LangChain Python cookbook for model discovery, chat, streaming, and tool calls through the standard text API.
• AIPowerGrid/n8n-nodes-aipg
  • dedicated n8n community node for text, image, video, and audio workflows with encrypted API-key credentials.
• starters/ - runnable on-chain game NPC, DAO media, Telegram, NFT media, and wallet-funded agent examples with quote-before-dispatch cost guards and no wallet keys in the inference process.

Weekly implementation evidence is recorded under proof/. Each report distinguishes local verification, upstream submission, upstream merge, and remaining release gates.

Builder-facing setup instructions and publication status are live at aipowergrid.io/docs/integrations. The public page treats compatibility, first-party tests, package publication, and upstream acceptance as separate states.

Builder credits

Builders shipping a public integration or demo can apply for $5-$20 of bounded service credit. Applications are public and manually reviewed; account linkage stays private. Review is not issuance: credits are granted only while the promotional spend rail is active, expire after 60 days, and cannot be transferred, withdrawn, or exchanged for cash or tokens. Requests may be processed by community-operated workers, so funded tests must use non-sensitive data.

Public conformance

Requires Node.js 20 or newer. The default run does not require a key and does not dispatch generation work:

npm run conformance:public

It checks service discovery, public modality and capacity metadata, and rejection of missing and invalid credentials before inference dispatch. The canonical OpenAI model-list shape and missing-model behavior are checked in authenticated account mode.

Account and quote conformance

Put a scoped API key in the environment, never in the command line:

GRID_API_KEY='...' node src/cli.mjs --account

This adds read-only credit-summary and canonical quote checks for each online modality. It does not submit a generation.

Explicit live text conformance

The live check spends against the authenticated account. It sends one bounded streaming request and validates event framing without saving its text:

GRID_API_KEY='...' node src/cli.mjs --account --live-text --model auto

The JSON report contains check names, outcomes, status codes, durations, and structural facts only. It excludes the key, prompts, generated content, account balances, response headers, and worker identity.

What this proves

A passing public run proves the currently deployed discovery and authentication boundary match this runner. An account run additionally proves non-mutating credit and quote contracts for the supplied account. A live-text run proves one small streaming request completed at that moment. None of these establish independent adoption, sustained uptime, private inference, or support for a third-party framework until that framework's own integration tests pass.

Verify everything

Install each child package with its locked package manager, then run:

npm run verify

The verification command is non-credentialed and does not submit generation jobs. Credentialed, credit-spending checks remain explicit release gates.

Generate an honest weekly proof snapshot

Generate the operational and payout copy directly from the public Grid APIs:

npm run proof:weekly

To refresh a dated artifact, pass an explicit output path. The generator reads only public network-status, job-total, and payout endpoints. It refuses an unknown status schema or economically active validator state, calculates job totals itself, and preserves paid-demand, independent-operator, external-builder, and historical-uptime metrics as explicit gaps when the public APIs cannot prove them.

Bounded production release gate

Before publishing a provider package, create a disposable key carrying only account.read and inference.submit, enable charging for that account, and run the four-integration gate:

AIPG_LIVE_E2E=1 AIPG_API_KEY='...' npm run release:e2e:live

The gate quotes six fixed requests before dispatching, requires sufficient credit and active charging, and refuses to exceed a hard $0.03 ceiling. It then exercises Dify, AI SDK, ElizaOS, and LangChain in sequence and checks that spend changed within the preflight bound. The n8n package owns its live release gate in its dedicated repository. Neither gate prints or persists the key, account balances, prompts, generated content, response headers, or worker identity. Revoke the disposable key immediately after the run.

After revoking it in the Console, keep the same environment variable only long enough to prove the credential is dead:

npm run release:key:verify-revoked
unset AIPG_API_KEY AIPG_LIVE_E2E AIPG_E2E_MAX_SPEND_USD

The revocation proof calls only the fixed production credit-summary endpoint, does not read a response body, and succeeds only on 401. Any other status is ambiguous and fails closed.

This package-level gate does not replace Dify Community Edition and Cloud UI installation checks, npm provenance, or upstream review.

Four integration and agent packages are public on npm:

• @aipowergrid/ai-sdk-provider@0.1.0
• @aipowergrid/plugin-aipg@0.1.0
• @aipowergrid/n8n-nodes-aipg@0.1.3
• @aipowergrid/mcp@0.2.0

Each package trusts only its matching GitHub Actions workflow through npm OIDC. The one-time bootstrap token and both repository NPM_TOKEN secrets were removed after first publication. Never publish locally or restore a long-lived registry token: the tag workflows are serialized, require the tag to match package.json, reject release commits that are not already on main, and produce npm provenance tied to the release commit.

The publication workflows deliberately disable setup-node package-manager caching. Release jobs carry OIDC permission, and restoring a writable package cache into that job creates an avoidable cache-poisoning path. Ordinary CI may still use lock-keyed caches because it has no package-publish permission.