refactor: generalize inference layer to multiple providers

.env.example

@@ -1,4 +1,19 @@
-OPENROUTER_API_KEY=your-key-here
+# Model provider API keys — set at least one. All providers are equal;
+# keys can also be added later in Admin → Model Providers.
+OPENROUTER_API_KEY=
+ANTHROPIC_API_KEY=
+OPENAI_API_KEY=
+GOOGLE_API_KEY=
+DEEPSEEK_API_KEY=
+GROQ_API_KEY=
+CEREBRAS_API_KEY=
+MISTRAL_API_KEY=
+XAI_API_KEY=
+ZAI_API_KEY=
+FIREWORKS_API_KEY=
+MINIMAX_API_KEY=
+HUGGINGFACE_API_KEY=
+
 BRAVE_SEARCH_API_KEY=your-key-here
 # 32+ char secrets. Generate each with: openssl rand -hex 32
 CREDENTIALS_KEY=

README.md

@@ -46,7 +46,7 @@ Self-hosted. Multi-user. Runs unattended.
 git clone https://github.com/0-AI-UG/zero-agent.git
 cd zero-agent && bun install
 
-cp .env.example .env   # add OPENROUTER_API_KEY, generate JWT_SECRET + CREDENTIALS_KEY
+cp .env.example .env   # add a model provider API key, generate JWT_SECRET + CREDENTIALS_KEY
 bun run dev
 ```
 
@@ -105,12 +105,12 @@ The same image backs `docker-compose.yml`, so `docker compose pull` fetches it i
 
 | Variable | Description |
 |---|---|
-| `OPENROUTER_API_KEY` | Required. [OpenRouter](https://openrouter.ai) key. |
+| `OPENROUTER_API_KEY` · `ANTHROPIC_API_KEY` · `OPENAI_API_KEY` · … | At least one model provider key is required; all providers are configured the same way (see `.env.example` for the full list). |
 | `JWT_SECRET` · `CREDENTIALS_KEY` | Required. ≥32 chars each (`openssl rand -hex 32`). |
 | `BRAVE_SEARCH_API_KEY` | Optional. Web search. |
 | `APP_URL` · `RP_ID` · `CORS_ORIGIN` · `TRUST_PROXY` | Set in production. |
 
-Models, image providers, and per-user limits are configured at runtime via the admin panel.
+Models, provider keys, capability routing (embeddings / image generation / captioning), and per-user limits are configured at runtime via the admin panel.
 
 ## Tech Stack
 

package.json

@@ -17,6 +17,10 @@
   "trustedDependencies": ["sharp", "bcrypt", "better-sqlite3", "esbuild"],
   "dependencies": {
     "@0-ai/s3lite": "^0.6.0",
+    "@ai-sdk/anthropic": "^3",
+    "@ai-sdk/google": "^3",
+    "@ai-sdk/openai": "^3",
+    "@ai-sdk/openai-compatible": "^2",
     "@anthropic-ai/sandbox-runtime": "^0.0.50",
     "@codemirror/lang-cpp": "^6.0.3",
     "@codemirror/lang-css": "^6.3.1",
@@ -30,14 +34,13 @@
     "@codemirror/lang-sql": "^6.10.0",
     "@codemirror/lang-xml": "^6.1.0",
     "@codemirror/lang-yaml": "^6.1.2",
-    "@hono/node-server": "^1.19.12",
     "@earendil-works/pi-agent-core": "^0.78.0",
     "@earendil-works/pi-ai": "^0.78.0",
     "@earendil-works/pi-coding-agent": "^0.78.0",
     "@earendil-works/pi-tui": "^0.78.0",
+    "@hono/node-server": "^1.19.12",
     "@mozilla/readability": "^0.6.0",
     "@openrouter/ai-sdk-provider": "^2.8.0",
-    "@openrouter/sdk": "^0.12.3",
     "@simplewebauthn/server": "^13.3.0",
     "@types/ws": "^8.18.1",
     "ai": "^6.0.164",

server/cli-handlers/embed.ts

@@ -4,8 +4,7 @@
  * embeddings without needing direct API key access.
  */
 import type { z } from "zod";
-import { embed } from "@/lib/openrouter/embed.ts";
-import { getEmbeddingModelId } from "@/lib/providers/index.ts";
+import { embed } from "@/lib/inference/embed.ts";
 import { isEmbeddingConfigured } from "@/lib/search/vectors.ts";
 import type { CliContext } from "./context.ts";
 import { ok, fail } from "./response.ts";
@@ -19,7 +18,7 @@ export async function handleEmbed(
     return fail("not_configured", "Embedding model is not configured", 503);
   }
 
-  const embeddings = await embed(input.texts, { model: getEmbeddingModelId() });
+  const embeddings = await embed(input.texts);
 
   return ok({
     embeddings,

server/cli-handlers/image.ts

@@ -1,11 +1,11 @@
 /**
- * Image generation handler — wraps generateImageViaOpenRouter, writes the
+ * Image generation handler — wraps generateImageViaProvider, writes the
  * bytes into the project directory so the agent can read them back, and
  * inserts a `files` row. The inotify watcher converges FTS / vectors
  * after the file lands.
  */
 import type { z } from "zod";
-import { generateImageViaOpenRouter } from "@/lib/media/image.ts";
+import { generateImageViaProvider } from "@/lib/media/image.ts";
 import { insertFile } from "@/db/queries/files.ts";
 import { createFolder as createFolderRecord, getFolderByPath } from "@/db/queries/folders.ts";
 import { writeProjectFile, workspacePathFor } from "@/lib/projects/fs-ops.ts";
@@ -31,7 +31,7 @@ export async function handleImageGenerate(
   ctx: CliContext,
   input: z.infer<typeof ImageGenerateInput>,
 ): Promise<Response> {
-  const image = await generateImageViaOpenRouter(input.prompt);
+  const image = await generateImageViaProvider(input.prompt);
   const timestamp = Date.now();
   const rawPath = input.path ?? `images/${timestamp}.png`;
   const filePath = sanitizePath(rawPath);

server/cli-handlers/llm.ts

@@ -4,8 +4,9 @@
  * without needing direct API key access.
  */
 import type { z } from "zod";
-import { generateText } from "@/lib/openrouter/text.ts";
+import { generateText } from "@/lib/inference/text.ts";
 import { getScriptsModelId } from "@/lib/providers/index.ts";
+import { resolveModelForPi } from "@/lib/pi/model.ts";
 import { insertUsageLog } from "@/db/queries/usage-logs.ts";
 import type { CliContext } from "./context.ts";
 import { ok } from "./response.ts";
@@ -16,10 +17,12 @@ export async function handleLlmGenerate(
   input: z.infer<typeof LlmGenerateInput>,
 ): Promise<Response> {
   const model = getScriptsModelId(ctx.projectId);
+  const resolved = resolveModelForPi(model);
 
   const result = await generateText({
-    model,
-    messages: input.prompt,
+    provider: resolved.provider,
+    model: resolved.modelId,
+    prompt: input.prompt,
     system: input.system,
     maxOutputTokens: input.maxTokens ?? 4096,
   });

server/db/queries/models.ts

@@ -49,7 +49,7 @@ export interface ModelInput {
   enabled?: boolean;
   sortOrder?: number;
   thinkingLevel?: ThinkingLevel | null;
-  piProvider?: string;
+  piProvider: string;
   piModelId?: string | null;
 }
 
@@ -64,7 +64,7 @@ export function insertModel(data: ModelInput): ModelRow {
     data.enabled !== false ? 1 : 0,
     data.sortOrder ?? 0,
     data.thinkingLevel ?? null,
-    data.piProvider ?? "openrouter",
+    data.piProvider,
     data.piModelId ?? null,
   );
   return getById.get(data.id) as ModelRow;

server/index.ts

@@ -167,6 +167,7 @@ import {
   handleAcceptUserInvitation,
 } from "@/routes/user-invitations.ts";
 import { handleSetupStatus, handleSetupComplete } from "@/routes/setup.ts";
+import { handleListProviders } from "@/routes/providers.ts";
 import { handleGetSettings, handleUpdateSettings, handleListImageModels } from "@/routes/settings.ts";
 import { handleEmailFeatureStatus, handleEmailFeatureToggle } from "@/routes/admin-email.ts";
 import {
@@ -428,6 +429,9 @@ app.post("/api/pending-responses/:id/respond", h(handleRespondPendingResponse));
 app.get("/api/setup/status", h(handleSetupStatus));
 app.post("/api/setup/complete", h(handleSetupComplete));
 
+// Model providers (no auth required — static, no secrets)
+app.get("/api/providers", h(handleListProviders));
+
 // Admin
 app.get("/api/admin/users", h(handleListUsers));
 app.post("/api/admin/users", h(handleCreateUser));

server/lib/chat-providers/email/router.ts

@@ -20,7 +20,7 @@ import { generateId, db } from "@/db/index.ts";
 import type { ChatRow, ProjectRow } from "@/db/types.ts";
 import { runTurn } from "@/lib/pi/run-turn.ts";
 import { resolveModelForPi } from "@/lib/pi/model.ts";
-import { getActiveProvider } from "@/lib/providers/index.ts";
+import { getDefaultChatModelId } from "@/lib/providers/index.ts";
 import { beginChatStream, endChatStream, publishPiEvent } from "@/lib/http/ws.ts";
 import { events as eventBus } from "@/lib/tasks/events.ts";
 
@@ -229,7 +229,7 @@ async function runEmailAgentTurn(
   const isFirstTurn = !existsSync(join(sessionsDirFor(project.id), `${chatId}.jsonl`));
   const priorOutbound = isFirstTurn ? outboundForChat(project.id, chatId) : [];
   const userText = composeUserMessage(parsed, attachments, priorOutbound, inboundEmailId);
-  const chatModelId = getActiveProvider().getDefaultChatModelId();
+  const chatModelId = getDefaultChatModelId();
   const resolved = resolveModelForPi(chatModelId);
 
   beginChatStream(chatId, "");