Intégrations externes : pouvoir créer des intégrations de type "fournisseur IA"

pierre-gilles · Août 1, 2026, 7:45

Pouvoir créer des intégrations externes qui puissent proposer des modèles d’IA

Demande basée sur : Sondage : Quel est votre modèle IA préféré dans Gladys? - #26 par l.dolle

pierre-gilles · Août 3, 2026, 2:52

La PR côté Gladys :

github.com/GladysAssistant/Gladys

Add generic AI provider external integration type (#2744)

master ← claude/generic-ai-provider-integration-oebx0g

ouvert 06:22AM - 02 Aug 26 UTC

Pierre-Gilles

+1021 -43

External integrations can now declare type "ai" and serve as the AI backend of t…he assistant instead of Gladys Plus. The contract is fully provider-agnostic: the whole agentic loop (MCP tool calling, prompts, history) stays in the core, and the integration only receives OpenAI-compatible chat completion requests over the integration WebSocket (external-integration.ai.chat, 120s ack) and answers with OpenAI-compatible completions — a thin adapter to Claude, DeepSeek, OpenAI, a local LLM or any other provider. Server: - manifest type "ai" (schema + validation), no device screens - ai.chat proxy service capability relayed by the supervisor - AI_PROVIDER system variable: gateway.aiChat routes all AI traffic (chat loop, intent router, weekly digest) to the selected provider, with the Gladys Plus model field stripped and no silent fallback - AI works without any Gladys Plus account (message.create) - GET/POST /api/v1/ai_provider endpoints, selection cleared on uninstall of the selected provider Front: - AI provider selection card on the AI integration page (admin) - generic external integration page branches for the ai type - install warning dedicated to AI providers, catalog placement - chat model selector hidden when an external provider is active - en/fr/de translations Spec: docs/specs/external-integrations.md B.18 + C.1/C.4/C.5/C.8 updated in the same diff (spec-first rule). Community request: https://community.gladysassistant.com/t/10421 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XhBhvoV4jdZEhxgadWwXV8 ## Summary by CodeRabbit * **New Features** * Added support for third-party AI providers using OpenAI-compatible chat completions. * Administrators can select, configure, retry, or reset the AI provider. * AI providers work without a Gladys Plus account. * Added provider availability indicators, privacy guidance, and dedicated setup flows. * **Bug Fixes** * Uninstalling the selected provider now restores the default AI service. * Improved installation routing and navigation for AI providers. * AI requests consistently use the selected provider and its configuration.

La PR côté SDK :

github.com/GladysAssistant/integration-sdk-js

feat: add AI provider support (manifest type "ai", onAiChat handler) (#22)

main ← claude/sdk-changes-pr-2744-d6ltvk

ouvert 03:46PM - 03 Aug 26 UTC

Pierre-Gilles

+275 -0

SDK counterpart of GladysAssistant/Gladys#2744 (generic AI provider external int…egration, contract B.19): integrations can now declare manifest `type: "ai"` and serve as the AI backend of Gladys in place of Gladys Plus. ## What's in - **New WebSocket message type** `external-integration.ai.chat` in `WEBSOCKET_MESSAGE_TYPES.EXTERNAL_INTEGRATION.AI_CHAT`. - **New `onAiChat(cb)` handler**: receives the OpenAI-compatible chat completion request relayed by the core (`{ messages, tools?, tool_choice?, purpose?, categories? }` — the `model` field is stripped core-side, picking one is the provider's job) and resolves an OpenAI-compatible completion, acked back as the `command-result` data. The core awaits the ack under **120 s** (`AI_CHAT_TIMEOUT_MS`), not the standard 5 s. - **Client-side completion validation**, mirroring the core's contract check: the resolved value must be an object carrying an object `choices[0].message` — anything else (including `undefined`) acks the command as failed with an explicit error, instead of letting a malformed completion travel to Gladys and die at the contract boundary (`ExternalIntegrationUnavailableError`). - **TypeScript typings**: `AiChatRequest`, `AiChatRequestMessage`, `AiChatCompletion`, `AiChatCompletionChoice`, `AiChatCompletionMessage`, the `onAiChat` method and the `AI_CHAT` message type. - **README**: new "AI providers" section (fetch-based OpenAI-compatible example, 120 s ack window, no-silent-fallback doctrine) + a row in the Handlers table. - **Tests**: `test/ai.test.js` against the fake Gladys server (success with `data`, tool_calls-only completion, handler throw, malformed completions, "not implemented"), plus compile-time coverage in `test/types/api.test-d.ts`. ## Checks - `npm test`: 185/185 pass - `npm run coverage`: thresholds green (100% stmts/funcs/lines) - `npm run lint`, `npm run prettier-check`, `npm run check-types`: clean 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_018mwqzunU7Hj5yXKoAwR1hi --- _Generated by [Claude Code](https://claude.ai/code/session_018mwqzunU7Hj5yXKoAwR1hi)_

cicoub13 · Octobre 3, 2026, 7:13

J’ai ajouté un commentaire sur les spécifications de la PR.

cursor-bot · Octobre 16, 2026, 10:08

Salut tout le monde !

Ce sujet est désormais en cours de développement.

Une PR a été ouverte pour ajouter le type d’intégration externe générique « ai » (fournisseurs d’IA) :

github.com/GladysAssistant/Gladys

Add the generic "ai" external integration type (AI providers) (#3211)

master ← claude/vigilant-cori-fjjpno

ouvert 10:07AM - 16 Oct 26 UTC

Pierre-Gilles

+1463 -48

### Description Supersedes #2744 (same design, rebased on the current master an…d reworked after @cicoub13's review there). External integrations can declare `type: "ai"` and serve as the AI backend of the assistant instead of Gladys Plus. The contract is provider-agnostic: the whole agentic loop (MCP tool calling, prompts, history) stays in the core, the integration only receives OpenAI-compatible chat completion requests over the integration WebSocket (`external-integration.ai.chat`, 120 s ack) and answers with OpenAI-compatible completions — a thin adapter to Claude, DeepSeek, OpenAI, a host Ollama or any other provider. **Server** - manifest type `ai` (validator + vendored schema), `ai.chat` proxy capability, `AI_PROVIDER` system variable, `GET`/`POST /api/v1/ai_provider`, selection cleared when the selected provider is uninstalled - `gateway.aiChat` routes all AI traffic (chat loop, intent router, scene `ai.ask`, weekly digest) to the selected provider, `model` stripped, no silent fallback to Gladys Plus - the completion is bounded before it enters the loop (content length, tool-call count, function names and arguments size, unknown fields dropped), and a cap of 4 requests in flight per provider protects local CPU models - `message.create` and `sendWeeklyDigest` work without Gladys Plus when an external provider is selected - the scene `ai.ask` action aborts with `AI_REQUEST_FAILED` on a failed request instead of a `TypeError` **Front** - `ai` joins `TYPES_WITHOUT_DEVICE_SCREENS` (install redirects, tabs, direct URLs), admin-only in the catalog, trust warning on the install screen - AI provider selection card on the AI integration page (trust warning restated, "provider not running" warning); Plus upsell, quota and rate-limit note hidden while an external provider is active, digest settings and debug download offered as soon as the assistant has a backend; chat model selector and scene "Ask AI" upsell hidden too - en / fr / de / es translations **Spec** - new `docs/specs/external-integrations/capabilities/ai-type.md` owning the type and its contracts (WS message, management routes, SDK `onAiChat`), per the post-split layout; `contracts/manifest.md`, `contracts/management-api.md`, `capabilities/provider-type.md` and `integration-catalog-categories.md` updated in the same diff **What changed compared with #2744, following the review there** 1. Spec moved from the deleted single file to `capabilities/ai-type.md` (no B.19 identifier, cited by path); `provider-type.md` records that `ai` passes the "generic pivot interface" bar. 2. The external-provider branch of `gateway.aiChat` sits before `callPlanGatedApi`, with a test pinning that a paused Plus subscription still reaches the provider. 3. `ai` added to `TYPES_WITHOUT_DEVICE_SCREENS`; catalog placement now comes from `categories` (`assistants`), not from the type. 4. Bundled-Ollama paragraph replaced by the documented host setup (`OLLAMA_HOST=172.30.0.1:11434`, context window, no NVIDIA grant to integrations); the voice sentence reduced to "speech is out of scope". 5. Weekly digest and `AskAI.jsx` gaps closed; completion bounded; `ai.ask` null path fixed; concurrency cap added. 6. `purpose` values documented as the real `AI_CHAT_PURPOSES` (`chat`, `intent-classification`, `weekly-digest`). Open points (recorded as "later leads" in the spec, not implemented here): a `provider` per `ai.ask` action, and selecting the integration at install time when no other provider is available. The companion SDK PR (GladysAssistant/integration-sdk-js#22) and the indexer enum still have to be updated in their own repos; until then `ai` integrations install through "Install from GitHub" and dev mode. ### Related request Forum: https://community.gladysassistant.com/t/10421 ### Checklist - [x] If a forum topic or GitHub issue exists, the description links it (`Forum: https://community.gladysassistant.com/t/...` or `Closes #...`) - [x] Tests pass: `cd server && npm run coverage` (Codecov requires 100% coverage on changed lines) and Cypress (`npm run cypress:run`) if the UI changed - [x] Linter and prettier pass on both front and server (`npm run eslint`, `npm run prettier`) - [x] No undocumented breaking change 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01JaNcYJ3imJoovepcWsxFSj --- _Generated by [Claude Code](https://claude.ai/code/session_01JaNcYJ3imJoovepcWsxFSj)_

N’hésitez pas à suivre la PR, à tester (optionnel, surtout pour les petites demandes) et à faire vos retours ici si besoin.

Sujet		Réponses	Vues
Intégrations externes : Pouvoir créer des intégrations de type "calendrier" Demandes livrées in-progress	1	68	Septembre 28, 2026
Filtrer la page des intégrations externes Demandes livrées	22	269	Octobre 23, 2026
Pouvoir configurer la conservation ou non des appareils des intégrations externes Demandes livrées feature-core , in-progress	1	57	Août 18, 2026
Intégration externe - Accéder au changelog de la version actuelle ou future Demandes livrées in-progress	1	51	Août 17, 2026
[Appareils] Avoir le vrai nom des intégrations Demande de fonctionnalités in-progress	6	89	Août 18, 2026

Intégrations externes : pouvoir créer des intégrations de type "fournisseur IA"

Sujets connexes