diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 8ed82589..b47837ca 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -9,6 +9,7 @@ jobs:
     runs-on: ubuntu-latest
     permissions:
       contents: write
+      packages: write
 
     steps:
       - uses: actions/checkout@v4
@@ -63,3 +64,57 @@ jobs:
           body_path: notes.md
           prerelease: ${{ steps.version.outputs.prerelease == 'true' }}
           generate_release_notes: false
+
+  publish-opentypebb:
+    runs-on: ubuntu-latest
+    needs: release
+    if: needs.release.result == 'success'
+    permissions:
+      contents: read
+      packages: write
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: pnpm/action-setup@v4
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          registry-url: https://npm.pkg.github.com
+          cache: pnpm
+
+      - name: Check if version already published
+        id: check
+        working-directory: packages/opentypebb
+        run: |
+          VERSION=$(node -p "require('./package.json').version")
+          echo "version=$VERSION" >> "$GITHUB_OUTPUT"
+          if npm view "@traderalice/opentypebb@$VERSION" version --registry=https://npm.pkg.github.com 2>/dev/null; then
+            echo "exists=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "exists=false" >> "$GITHUB_OUTPUT"
+          fi
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Build
+        if: steps.check.outputs.exists == 'false'
+        working-directory: packages/opentypebb
+        run: |
+          pnpm install --frozen-lockfile
+          pnpm build
+
+      - name: Publish to GitHub Packages
+        if: steps.check.outputs.exists == 'false'
+        working-directory: packages/opentypebb
+        run: npm publish
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Publish to npmjs
+        if: steps.check.outputs.exists == 'false'
+        working-directory: packages/opentypebb
+        run: npm publish --registry=https://registry.npmjs.org
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
diff --git a/src/skills/createDocx.md b/data/default/skills/createDocx.md
similarity index 100%
rename from src/skills/createDocx.md
rename to data/default/skills/createDocx.md
diff --git a/package.json b/package.json
index c303b77c..3db2f6b1 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "open-alice",
-  "version": "0.9.0-beta.1",
+  "version": "0.9.0-beta.2",
   "description": "File-based trading agent engine",
   "type": "module",
   "scripts": {
@@ -47,7 +47,7 @@
     "grammy": "^1.40.0",
     "hono": "^4.12.5",
     "json5": "^2.2.3",
-    "opentypebb": "link:./packages/opentypebb",
+    "@traderalice/opentypebb": "link:./packages/opentypebb",
     "pino": "^10.3.1",
     "playwright-core": "1.58.2",
     "sharp": "^0.34.5",
diff --git a/packages/opentypebb/package.json b/packages/opentypebb/package.json
index a9155d93..85d6804f 100644
--- a/packages/opentypebb/package.json
+++ b/packages/opentypebb/package.json
@@ -1,5 +1,5 @@
 {
-  "name": "opentypebb",
+  "name": "@traderalice/opentypebb",
   "version": "0.1.0",
   "description": "TypeScript port of OpenBB Platform — financial data infrastructure",
   "type": "module",
@@ -13,6 +13,27 @@
       "types": "./src/server.ts"
     }
   },
+  "files": [
+    "dist"
+  ],
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/TraderAlice/OpenAlice.git",
+    "directory": "packages/opentypebb"
+  },
+  "publishConfig": {
+    "registry": "https://npm.pkg.github.com",
+    "exports": {
+      ".": {
+        "import": "./dist/index.js",
+        "types": "./dist/index.d.ts"
+      },
+      "./server": {
+        "import": "./dist/server.js",
+        "types": "./dist/server.d.ts"
+      }
+    }
+  },
   "scripts": {
     "dev": "tsx src/server.ts",
     "build": "tsup",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 773857fa..0b0de3db 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -35,6 +35,9 @@ importers:
       '@sinclair/typebox':
         specifier: 0.34.48
         version: 0.34.48
+      '@traderalice/opentypebb':
+        specifier: link:./packages/opentypebb
+        version: link:packages/opentypebb
       ai:
         specifier: ^6.0.86
         version: 6.0.86(zod@4.3.6)
@@ -65,9 +68,6 @@ importers:
       json5:
         specifier: ^2.2.3
         version: 2.2.3
-      opentypebb:
-        specifier: link:./packages/opentypebb
-        version: link:packages/opentypebb
       pino:
         specifier: ^10.3.1
         version: 10.3.1
diff --git a/src/ai-providers/agent-sdk/agent-sdk-provider.ts b/src/ai-providers/agent-sdk/agent-sdk-provider.ts
index a9322f2a..689c3539 100644
--- a/src/ai-providers/agent-sdk/agent-sdk-provider.ts
+++ b/src/ai-providers/agent-sdk/agent-sdk-provider.ts
@@ -10,8 +10,7 @@
 
 import { resolve } from 'node:path'
 import type { Tool } from 'ai'
-import type { ProviderResult, ProviderEvent } from '../../core/ai-provider.js'
-import type { GenerateProvider, GenerateInput, GenerateOpts } from '../../core/ai-provider.js'
+import type { ProviderResult, ProviderEvent, GenerateProvider, GenerateInput, GenerateOpts } from '../types.js'
 import type { AgentSdkConfig, AgentSdkOverride } from './query.js'
 import { readAgentConfig } from '../../core/config.js'
 import { extractMediaFromToolResultContent } from '../../core/media.js'
diff --git a/src/ai-providers/agent-sdk/tool-bridge.ts b/src/ai-providers/agent-sdk/tool-bridge.ts
index ebe898d1..9b424801 100644
--- a/src/ai-providers/agent-sdk/tool-bridge.ts
+++ b/src/ai-providers/agent-sdk/tool-bridge.ts
@@ -1,7 +1,7 @@
 /**
  * Tool bridge — converts ToolCenter's Vercel AI SDK tools to an Agent SDK MCP server.
  *
- * Reuses the same pattern as `src/plugins/mcp.ts` (extract .shape, wrap execute),
+ * Reuses the same pattern as `src/server/mcp.ts` (extract .shape, wrap execute),
  * but targets `createSdkMcpServer()` instead of `@modelcontextprotocol/sdk McpServer`.
  */
 
diff --git a/src/ai-providers/claude-code/claude-code-provider.ts b/src/ai-providers/claude-code/claude-code-provider.ts
index e97b4e82..5ce5482b 100644
--- a/src/ai-providers/claude-code/claude-code-provider.ts
+++ b/src/ai-providers/claude-code/claude-code-provider.ts
@@ -9,8 +9,7 @@
  */
 
 import { resolve } from 'node:path'
-import type { ProviderResult, ProviderEvent } from '../../core/ai-provider.js'
-import type { GenerateProvider, GenerateInput, GenerateOpts } from '../../core/ai-provider.js'
+import type { ProviderResult, ProviderEvent, GenerateProvider, GenerateInput, GenerateOpts } from '../types.js'
 import type { ClaudeCodeConfig } from './types.js'
 import { readAgentConfig } from '../../core/config.js'
 import { extractMediaFromToolResultContent } from '../../core/media.js'
diff --git a/src/ai-providers/types.ts b/src/ai-providers/types.ts
new file mode 100644
index 00000000..c18bc9d8
--- /dev/null
+++ b/src/ai-providers/types.ts
@@ -0,0 +1,64 @@
+import type { SessionStore, SDKModelMessage } from '../core/session.js'
+import type { CompactionConfig, CompactionResult } from '../core/compaction.js'
+import type { MediaAttachment } from '../core/types.js'
+
+// ==================== Provider Events ====================
+
+/** Streaming event emitted by AI providers during generation. */
+export type ProviderEvent =
+  | { type: 'tool_use'; id: string; name: string; input: unknown }
+  | { type: 'tool_result'; tool_use_id: string; content: string }
+  | { type: 'text'; text: string }
+  | { type: 'done'; result: ProviderResult }
+
+// ==================== Types ====================
+
+export interface ProviderResult {
+  text: string
+  media: MediaAttachment[]
+  mediaUrls?: string[]
+}
+
+// ==================== GenerateProvider ====================
+
+/**
+ * Input prepared by AgentCenter, dispatched by provider.inputKind.
+ *
+ * - 'text': Claude Code / Agent SDK — single string prompt with <chat_history> baked in.
+ * - 'messages': Vercel AI SDK — structured ModelMessage[] (history carried natively).
+ */
+export type GenerateInput =
+  | { kind: 'text'; prompt: string; systemPrompt?: string }
+  | { kind: 'messages'; messages: SDKModelMessage[]; systemPrompt?: string }
+
+/** Per-request options passed through to the underlying provider. */
+export interface GenerateOpts {
+  disabledTools?: string[]
+  vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
+  agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
+}
+
+/**
+ * Slim provider interface — pure data-source adapter.
+ *
+ * Does NOT touch session management. AgentCenter prepares the input,
+ * the provider calls the backend and yields ProviderEvents.
+ */
+export interface GenerateProvider {
+  /** Which input format this provider expects. */
+  readonly inputKind: 'text' | 'messages'
+  /** Session log provenance tag. */
+  readonly providerTag: 'vercel-ai' | 'claude-code' | 'agent-sdk'
+  /** Stateless one-shot prompt (used for compaction summarization, etc.). */
+  ask(prompt: string): Promise<ProviderResult>
+  /** Stream events from the backend. Yields tool_use/tool_result/text, then done. */
+  generate(input: GenerateInput, opts?: GenerateOpts): AsyncIterable<ProviderEvent>
+  /**
+   * Optional: custom compaction strategy. If implemented, AgentCenter delegates
+   * compaction to the provider instead of using the default compactIfNeeded.
+   *
+   * Use case: providers with native server-side compaction (e.g. Anthropic API
+   * compact-2026-01-12) can bypass the local JSONL-based summarization.
+   */
+  compact?(session: SessionStore, config: CompactionConfig): Promise<CompactionResult>
+}
diff --git a/src/core/model-factory.ts b/src/ai-providers/vercel-ai-sdk/model-factory.ts
similarity index 97%
rename from src/core/model-factory.ts
rename to src/ai-providers/vercel-ai-sdk/model-factory.ts
index 8326bf85..447cc0eb 100644
--- a/src/core/model-factory.ts
+++ b/src/ai-providers/vercel-ai-sdk/model-factory.ts
@@ -7,7 +7,7 @@
  */
 
 import type { LanguageModel } from 'ai'
-import { readAIProviderConfig } from './config.js'
+import { readAIProviderConfig } from '../../core/config.js'
 
 /** Result includes the model plus a cache key for change detection. */
 export interface ModelFromConfig {
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
index 539eee2c..31e09190 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
@@ -7,12 +7,11 @@
  */
 
 import type { ModelMessage, Tool } from 'ai'
-import type { ProviderResult, ProviderEvent } from '../../core/ai-provider.js'
-import type { GenerateProvider, GenerateInput, GenerateOpts } from '../../core/ai-provider.js'
+import type { ProviderResult, ProviderEvent, GenerateProvider, GenerateInput, GenerateOpts } from '../types.js'
 import type { Agent } from './agent.js'
 import type { MediaAttachment } from '../../core/types.js'
 import { extractMediaFromToolOutput } from '../../core/media.js'
-import { createModelFromConfig, type ModelOverride } from '../../core/model-factory.js'
+import { createModelFromConfig, type ModelOverride } from './model-factory.js'
 import { createAgent } from './agent.js'
 import { createChannel } from '../../core/async-channel.js'
 
diff --git a/src/connectors/telegram/telegram-plugin.ts b/src/connectors/telegram/telegram-plugin.ts
index fa5a2811..fdb6b757 100644
--- a/src/connectors/telegram/telegram-plugin.ts
+++ b/src/connectors/telegram/telegram-plugin.ts
@@ -10,8 +10,9 @@ import { askClaudeCode } from '../../ai-providers/claude-code/index.js'
 import type { ClaudeCodeConfig } from '../../ai-providers/claude-code/index.js'
 import { SessionStore } from '../../core/session'
 import { forceCompact } from '../../core/compaction'
-import { readAIConfig, writeAIConfig, type AIBackend } from '../../core/ai-config'
-import type { ConnectorCenter, Connector } from '../../core/connector-center.js'
+import { readAIBackend, writeAIBackend, type AIBackend } from '../../core/config'
+import type { ConnectorCenter } from '../../core/connector-center.js'
+import type { Connector } from '../types.js'
 
 const MAX_MESSAGE_LENGTH = 4096
 
@@ -82,7 +83,7 @@ export class TelegramPlugin implements Plugin {
 
     // ── Commands ──
     bot.command('status', async (ctx) => {
-      const aiConfig = await readAIConfig()
+      const aiConfig = await readAIBackend()
       await this.sendReply(ctx.chat.id, `Engine is running. Provider: ${BACKEND_LABELS[aiConfig.backend]}`)
     })
 
@@ -106,7 +107,7 @@ export class TelegramPlugin implements Plugin {
       try {
         if (data.startsWith('provider:')) {
           const backend = data.slice('provider:'.length) as AIBackend
-          await writeAIConfig(backend)
+          await writeAIBackend(backend)
           await ctx.answerCallbackQuery({ text: `Switched to ${BACKEND_LABELS[backend]}` })
 
           // Edit the original settings message in-place
@@ -170,7 +171,7 @@ export class TelegramPlugin implements Plugin {
 
     // ── Initialize and get bot info ──
     await bot.init()
-    const aiConfig = await readAIConfig()
+    const aiConfig = await readAIBackend()
     console.log(`telegram plugin: connected as @${bot.botInfo.username} (backend: ${aiConfig.backend})`)
 
     // ── Register connector for outbound delivery (heartbeat / cron responses) ──
@@ -320,7 +321,7 @@ export class TelegramPlugin implements Plugin {
   }
 
   private async sendSettingsMenu(chatId: number) {
-    const aiConfig = await readAIConfig()
+    const aiConfig = await readAIBackend()
     const ccLabel = aiConfig.backend === 'claude-code' ? '> Claude Code' : 'Claude Code'
     const aiLabel = aiConfig.backend === 'vercel-ai-sdk' ? '> Vercel AI SDK' : 'Vercel AI SDK'
     const sdkLabel = aiConfig.backend === 'agent-sdk' ? '> Agent SDK' : 'Agent SDK'
diff --git a/src/connectors/types.ts b/src/connectors/types.ts
new file mode 100644
index 00000000..9482902a
--- /dev/null
+++ b/src/connectors/types.ts
@@ -0,0 +1,58 @@
+import type { MediaAttachment } from '../core/types.js'
+import type { StreamableResult } from '../core/ai-provider.js'
+
+// ==================== Send Types ====================
+
+/** Structured payload for outbound send (heartbeat, cron, manual, etc.). */
+export interface SendPayload {
+  /** Whether this is a chat message or a notification. */
+  kind: 'message' | 'notification'
+  /** The text content to send. */
+  text: string
+  /** Media attachments (e.g. screenshots from tools). */
+  media?: MediaAttachment[]
+  /** Where this payload originated from. */
+  source?: 'heartbeat' | 'cron' | 'manual'
+}
+
+/** Result of a send() call. */
+export interface SendResult {
+  /** Whether the message was actually sent (false for pull-based connectors). */
+  delivered: boolean
+}
+
+// ==================== Connector Interface ====================
+
+/** Discoverable capabilities a connector may support. */
+export interface ConnectorCapabilities {
+  /** Can push messages proactively (heartbeat/cron). False for pull-based. */
+  push: boolean
+  /** Can send media attachments (images). */
+  media: boolean
+}
+
+/**
+ * A connector that can send outbound messages to a user.
+ *
+ * Each plugin (Telegram, Web, MCP-ask) implements this interface and
+ * registers itself with the ConnectorCenter.
+ */
+export interface Connector {
+  /** Channel identifier, e.g. "telegram", "web", "mcp-ask". */
+  readonly channel: string
+  /** Recipient identifier (chat id, "default", session id, etc.). */
+  readonly to: string
+  /** What this connector can do. */
+  readonly capabilities: ConnectorCapabilities
+  /** Send a structured payload through this connector. */
+  send(payload: SendPayload): Promise<SendResult>
+  /**
+   * Optional: stream AI response events to the client in real-time.
+   * Connectors that support this can push ProviderEvents (tool_use, tool_result, text)
+   * as they arrive, then deliver the final result at the end.
+   *
+   * If not implemented, ConnectorCenter falls back to draining the stream
+   * and calling send() with the completed result.
+   */
+  sendStream?(stream: StreamableResult, meta?: Pick<SendPayload, 'kind' | 'source'>): Promise<SendResult>
+}
diff --git a/src/connectors/web/routes/config.ts b/src/connectors/web/routes/config.ts
index 83688aee..6e3c55d2 100644
--- a/src/connectors/web/routes/config.ts
+++ b/src/connectors/web/routes/config.ts
@@ -1,6 +1,5 @@
 import { Hono } from 'hono'
-import { loadConfig, writeConfigSection, readAIProviderConfig, readOpenbbConfig, validSections, type ConfigSection } from '../../../core/config.js'
-import { readAIConfig, writeAIConfig, type AIBackend } from '../../../core/ai-config.js'
+import { loadConfig, writeConfigSection, readAIProviderConfig, readOpenbbConfig, validSections, writeAIBackend, type ConfigSection, type AIBackend } from '../../../core/config.js'
 
 interface ConfigRouteOpts {
   onConnectorsChange?: () => Promise<void>
@@ -26,7 +25,7 @@ export function createConfigRoutes(opts?: ConfigRouteOpts) {
       if (backend !== 'claude-code' && backend !== 'vercel-ai-sdk' && backend !== 'agent-sdk') {
         return c.json({ error: 'Invalid backend. Must be "claude-code", "vercel-ai-sdk", or "agent-sdk".' }, 400)
       }
-      await writeAIConfig(backend as AIBackend)
+      await writeAIBackend(backend as AIBackend)
       return c.json({ backend })
     } catch (err) {
       return c.json({ error: String(err) }, 500)
diff --git a/src/connectors/web/web-plugin.ts b/src/connectors/web/web-plugin.ts
index 3c1baec4..11b55435 100644
--- a/src/connectors/web/web-plugin.ts
+++ b/src/connectors/web/web-plugin.ts
@@ -5,7 +5,7 @@ import { serveStatic } from '@hono/node-server/serve-static'
 import { resolve } from 'node:path'
 import type { Plugin, EngineContext } from '../../core/types.js'
 import { SessionStore, type ContentBlock } from '../../core/session.js'
-import type { Connector, SendPayload } from '../../core/connector-center.js'
+import type { Connector, SendPayload } from '../types.js'
 import type { StreamableResult } from '../../core/ai-provider.js'
 import { persistMedia } from '../../core/media-store.js'
 import { readWebSubchannels } from '../../core/config.js'
diff --git a/src/core/__tests__/pipeline/persistence.spec.ts b/src/core/__tests__/pipeline/persistence.spec.ts
index 9bfddb91..396473c3 100644
--- a/src/core/__tests__/pipeline/persistence.spec.ts
+++ b/src/core/__tests__/pipeline/persistence.spec.ts
@@ -309,7 +309,7 @@ describe('AgentCenter — session persistence', () => {
     )
   })
 
-  it('A12: multiple consecutive text events all buffered in intermediate flush', async () => {
+  it('A12: multiple consecutive text events persisted once via final result', async () => {
     const provider = new FakeProvider([
       textEvent('first '),
       textEvent('second '),
@@ -323,19 +323,9 @@ describe('AgentCenter — session persistence', () => {
 
     const assistantWrites = session.writes.filter(w => w.method === 'appendAssistant')
 
-    const intermediateFlush = assistantWrites.find(w => {
-      const content = w.content as ContentBlock[]
-      return Array.isArray(content) && content.filter(b => b.type === 'text').length === 3
-    })
-    expect(intermediateFlush).toBeDefined()
-    expect(intermediateFlush!.content).toEqual([
-      { type: 'text', text: 'first ' },
-      { type: 'text', text: 'second ' },
-      { type: 'text', text: 'third' },
-    ])
-
-    const finalWrite = assistantWrites[assistantWrites.length - 1]
-    expect(finalWrite.content).toEqual([{ type: 'text', text: 'first second third' }])
+    // Only one assistant write — the authoritative final text (no duplicate intermediate flush)
+    expect(assistantWrites).toHaveLength(1)
+    expect(assistantWrites[0].content).toEqual([{ type: 'text', text: 'first second third' }])
   })
 
   it('A13: tool_use with complex nested input preserves structure', async () => {
diff --git a/src/core/engine.spec.ts b/src/core/agent-center.spec.ts
similarity index 97%
rename from src/core/engine.spec.ts
rename to src/core/agent-center.spec.ts
index 92bf0d97..7fb28870 100644
--- a/src/core/engine.spec.ts
+++ b/src/core/agent-center.spec.ts
@@ -5,7 +5,7 @@ import { AgentCenter } from './agent-center.js'
 import { GenerateRouter } from './ai-provider.js'
 import { DEFAULT_COMPACTION_CONFIG, type CompactionConfig } from './compaction.js'
 import { VercelAIProvider } from '../ai-providers/vercel-ai-sdk/vercel-provider.js'
-import { createModelFromConfig } from './model-factory.js'
+import { createModelFromConfig } from '../ai-providers/vercel-ai-sdk/model-factory.js'
 import type { SessionStore, SessionEntry } from './session.js'
 
 // ==================== Helpers ====================
@@ -43,7 +43,7 @@ function makeAgentCenter(overrides: MakeAgentCenterOpts = {}): AgentCenter {
   const compaction = overrides.compaction ?? DEFAULT_COMPACTION_CONFIG
 
   vi.mocked(createModelFromConfig).mockResolvedValue({ model, key: 'test:mock-model' })
-  const provider = new VercelAIProvider(() => tools, instructions, maxSteps)
+  const provider = new VercelAIProvider(async () => tools, instructions, maxSteps)
   const router = new GenerateRouter(provider, null)
 
   return new AgentCenter({ router, compaction })
@@ -88,7 +88,7 @@ function makeSessionMock(entries: SessionEntry[] = []): SessionStore {
 
 // ==================== Mock model-factory ====================
 
-vi.mock('./model-factory.js', () => ({
+vi.mock('../ai-providers/vercel-ai-sdk/model-factory.js', () => ({
   createModelFromConfig: vi.fn(),
 }))
 
diff --git a/src/core/agent-center.ts b/src/core/agent-center.ts
index ae4e2444..5f696fdc 100644
--- a/src/core/agent-center.ts
+++ b/src/core/agent-center.ts
@@ -165,10 +165,9 @@ export class AgentCenter {
       }
     }
 
-    // Flush any remaining intermediate blocks
-    if (currentAssistantBlocks.length > 0) {
-      intermediateMessages.push({ role: 'assistant', content: currentAssistantBlocks })
-    }
+    // Flush any remaining user blocks (defensive — tool_result already flushes)
+    // NOTE: Do NOT flush trailing assistant text blocks here — the authoritative
+    // final text comes from the done event and is persisted once in step 8.
     if (currentUserBlocks.length > 0) {
       intermediateMessages.push({ role: 'user', content: currentUserBlocks })
     }
diff --git a/src/core/ai-config.ts b/src/core/ai-config.ts
deleted file mode 100644
index d142b15e..00000000
--- a/src/core/ai-config.ts
+++ /dev/null
@@ -1,19 +0,0 @@
-import { readFile, writeFile, mkdir } from 'node:fs/promises'
-import { resolve } from 'node:path'
-import { readAIProviderConfig } from './config.js'
-
-export type AIBackend = 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk'
-
-const CONFIG_PATH = resolve('data/config/ai-provider.json')
-
-export async function readAIConfig() {
-  const config = await readAIProviderConfig()
-  return { backend: config.backend }
-}
-
-export async function writeAIConfig(backend: AIBackend): Promise<void> {
-  const current = await readAIProviderConfig()
-  const updated = { ...current, backend }
-  await mkdir(resolve('data/config'), { recursive: true })
-  await writeFile(CONFIG_PATH, JSON.stringify(updated, null, 2) + '\n')
-}
diff --git a/src/core/ai-provider.ts b/src/core/ai-provider.ts
index c5bde42e..b8f69911 100644
--- a/src/core/ai-provider.ts
+++ b/src/core/ai-provider.ts
@@ -1,27 +1,18 @@
 /**
- * AI Provider abstraction — GenerateProvider + GenerateRouter.
+ * AI Provider abstraction — StreamableResult + GenerateRouter.
  *
- * GenerateProvider is a slim data-source adapter: each backend (Vercel AI SDK,
- * Claude Code CLI, Agent SDK) implements `ask()` and `generate()`.
- * Session management lives in AgentCenter, not here.
- *
- * GenerateRouter reads runtime config and resolves to the correct provider.
+ * Provider interface types (GenerateProvider, ProviderEvent, etc.) live in
+ * ai-providers/types.ts alongside the implementations. This file holds the
+ * core infrastructure that orchestrates providers.
  */
 
-import type { SessionStore } from './session.js'
-import type { SDKModelMessage } from './session.js'
-import type { CompactionConfig, CompactionResult } from './compaction.js'
-import type { MediaAttachment } from './types.js'
 import { readAIProviderConfig } from './config.js'
+import type { ProviderEvent, ProviderResult, GenerateProvider } from '../ai-providers/types.js'
 
-// ==================== Provider Events ====================
-
-/** Streaming event emitted by AI providers during generation. */
-export type ProviderEvent =
-  | { type: 'tool_use'; id: string; name: string; input: unknown }
-  | { type: 'tool_result'; tool_use_id: string; content: string }
-  | { type: 'text'; text: string }
-  | { type: 'done'; result: ProviderResult }
+export type {
+  ProviderEvent, ProviderResult, GenerateProvider,
+  GenerateInput, GenerateOpts,
+} from '../ai-providers/types.js'
 
 // ==================== StreamableResult ====================
 
@@ -140,56 +131,6 @@ export interface AskOptions {
   }
 }
 
-export interface ProviderResult {
-  text: string
-  media: MediaAttachment[]
-  mediaUrls?: string[]
-}
-
-// ==================== GenerateProvider ====================
-
-/**
- * Input prepared by AgentCenter, dispatched by provider.inputKind.
- *
- * - 'text': Claude Code / Agent SDK — single string prompt with <chat_history> baked in.
- * - 'messages': Vercel AI SDK — structured ModelMessage[] (history carried natively).
- */
-export type GenerateInput =
-  | { kind: 'text'; prompt: string; systemPrompt?: string }
-  | { kind: 'messages'; messages: SDKModelMessage[]; systemPrompt?: string }
-
-/** Per-request options passed through to the underlying provider. */
-export interface GenerateOpts {
-  disabledTools?: string[]
-  vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
-  agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
-}
-
-/**
- * Slim provider interface — pure data-source adapter.
- *
- * Does NOT touch session management. AgentCenter prepares the input,
- * the provider calls the backend and yields ProviderEvents.
- */
-export interface GenerateProvider {
-  /** Which input format this provider expects. */
-  readonly inputKind: 'text' | 'messages'
-  /** Session log provenance tag. */
-  readonly providerTag: 'vercel-ai' | 'claude-code' | 'agent-sdk'
-  /** Stateless one-shot prompt (used for compaction summarization, etc.). */
-  ask(prompt: string): Promise<ProviderResult>
-  /** Stream events from the backend. Yields tool_use/tool_result/text, then done. */
-  generate(input: GenerateInput, opts?: GenerateOpts): AsyncIterable<ProviderEvent>
-  /**
-   * Optional: custom compaction strategy. If implemented, AgentCenter delegates
-   * compaction to the provider instead of using the default compactIfNeeded.
-   *
-   * Use case: providers with native server-side compaction (e.g. Anthropic API
-   * compact-2026-01-12) can bypass the local JSONL-based summarization.
-   */
-  compact?(session: SessionStore, config: CompactionConfig): Promise<CompactionResult>
-}
-
 // ==================== GenerateRouter ====================
 
 /** Reads runtime AI config and resolves to the correct GenerateProvider. */
diff --git a/src/core/config.ts b/src/core/config.ts
index 7fa71fbe..a8605b7f 100644
--- a/src/core/config.ts
+++ b/src/core/config.ts
@@ -518,6 +518,24 @@ export async function readToolsConfig() {
   }
 }
 
+// ==================== AI Backend Helpers ====================
+
+export type AIBackend = 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk'
+
+/** Read the current AI backend from ai-provider.json. */
+export async function readAIBackend(): Promise<{ backend: AIBackend }> {
+  const config = await readAIProviderConfig()
+  return { backend: config.backend }
+}
+
+/** Switch the AI backend in ai-provider.json (preserves other fields). */
+export async function writeAIBackend(backend: AIBackend): Promise<void> {
+  const current = await readAIProviderConfig()
+  const updated = { ...current, backend }
+  await mkdir(CONFIG_DIR, { recursive: true })
+  await writeFile(resolve(CONFIG_DIR, 'ai-provider.json'), JSON.stringify(updated, null, 2) + '\n')
+}
+
 // ==================== Writer ====================
 
 export type ConfigSection = keyof Config
diff --git a/src/core/connector-center.ts b/src/core/connector-center.ts
index 0df07f37..241c7638 100644
--- a/src/core/connector-center.ts
+++ b/src/core/connector-center.ts
@@ -10,65 +10,12 @@
  * replies go to whichever channel the user most recently interacted through.
  */
 
-import type { MediaAttachment } from './types.js'
 import type { EventLog } from './event-log.js'
+import type { MediaAttachment } from './types.js'
 import type { StreamableResult } from './ai-provider.js'
+import type { Connector, SendPayload, SendResult } from '../connectors/types.js'
 
-// ==================== Send Types ====================
-
-/** Structured payload for outbound send (heartbeat, cron, manual, etc.). */
-export interface SendPayload {
-  /** Whether this is a chat message or a notification. */
-  kind: 'message' | 'notification'
-  /** The text content to send. */
-  text: string
-  /** Media attachments (e.g. screenshots from tools). */
-  media?: MediaAttachment[]
-  /** Where this payload originated from. */
-  source?: 'heartbeat' | 'cron' | 'manual'
-}
-
-/** Result of a send() call. */
-export interface SendResult {
-  /** Whether the message was actually sent (false for pull-based connectors). */
-  delivered: boolean
-}
-
-// ==================== Connector Interface ====================
-
-/** Discoverable capabilities a connector may support. */
-export interface ConnectorCapabilities {
-  /** Can push messages proactively (heartbeat/cron). False for pull-based. */
-  push: boolean
-  /** Can send media attachments (images). */
-  media: boolean
-}
-
-/**
- * A connector that can send outbound messages to a user.
- *
- * Each plugin (Telegram, Web, MCP-ask) implements this interface and
- * registers itself with the ConnectorCenter.
- */
-export interface Connector {
-  /** Channel identifier, e.g. "telegram", "web", "mcp-ask". */
-  readonly channel: string
-  /** Recipient identifier (chat id, "default", session id, etc.). */
-  readonly to: string
-  /** What this connector can do. */
-  readonly capabilities: ConnectorCapabilities
-  /** Send a structured payload through this connector. */
-  send(payload: SendPayload): Promise<SendResult>
-  /**
-   * Optional: stream AI response events to the client in real-time.
-   * Connectors that support this can push ProviderEvents (tool_use, tool_result, text)
-   * as they arrive, then deliver the final result at the end.
-   *
-   * If not implemented, ConnectorCenter falls back to draining the stream
-   * and calling send() with the completed result.
-   */
-  sendStream?(stream: StreamableResult, meta?: Pick<SendPayload, 'kind' | 'source'>): Promise<SendResult>
-}
+export type { Connector, SendPayload, SendResult, ConnectorCapabilities } from '../connectors/types.js'
 
 // ==================== Notify Types ====================
 
diff --git a/src/main.ts b/src/main.ts
index 169ffb24..ceae7be4 100644
--- a/src/main.ts
+++ b/src/main.ts
@@ -3,7 +3,7 @@ import { resolve, dirname } from 'path'
 // Engine removed — AgentCenter is the top-level AI entry point
 import { loadConfig, loadTradingConfig } from './core/config.js'
 import type { Plugin, EngineContext, ReconnectResult } from './core/types.js'
-import { McpPlugin } from './plugins/mcp.js'
+import { McpPlugin } from './server/mcp.js'
 import { TelegramPlugin } from './connectors/telegram/index.js'
 import { WebPlugin } from './connectors/web/index.js'
 import { McpAskPlugin } from './connectors/mcp-ask/index.js'
@@ -31,7 +31,7 @@ import { OpenBBEquityClient } from './openbb/equity/client.js'
 import { OpenBBCryptoClient } from './openbb/crypto/client.js'
 import { OpenBBCurrencyClient } from './openbb/currency/client.js'
 import { OpenBBNewsClient } from './openbb/news/client.js'
-import { startEmbeddedOpenBBServer } from './openbb/api-server.js'
+import { startEmbeddedOpenBBServer } from './server/opentypebb.js'
 import { createMarketSearchTools } from './extension/market/index.js'
 import { createNewsTools } from './extension/news/index.js'
 import { createAnalysisTools } from './extension/analysis-kit/index.js'
diff --git a/src/openbb/sdk/base-client.ts b/src/openbb/sdk/base-client.ts
index 1d717bab..fbfea082 100644
--- a/src/openbb/sdk/base-client.ts
+++ b/src/openbb/sdk/base-client.ts
@@ -11,7 +11,7 @@
  *     → executor.execute('fmp', 'EquityQuote', params, credentials)
  */
 
-import type { QueryExecutor } from 'opentypebb'
+import type { QueryExecutor } from '@traderalice/opentypebb'
 
 export class SDKBaseClient {
   constructor(
diff --git a/src/openbb/sdk/executor.ts b/src/openbb/sdk/executor.ts
index 8b87cc97..18aadc12 100644
--- a/src/openbb/sdk/executor.ts
+++ b/src/openbb/sdk/executor.ts
@@ -6,7 +6,7 @@
  * without HTTP overhead.
  */
 
-import { createExecutor, type QueryExecutor } from 'opentypebb'
+import { createExecutor, type QueryExecutor } from '@traderalice/opentypebb'
 
 let _executor: QueryExecutor | null = null
 
diff --git a/src/openbb/sdk/route-map.ts b/src/openbb/sdk/route-map.ts
index fa4615df..c04a2543 100644
--- a/src/openbb/sdk/route-map.ts
+++ b/src/openbb/sdk/route-map.ts
@@ -8,7 +8,7 @@
  * to call for each API path, providing a drop-in replacement for HTTP routing.
  */
 
-import { loadAllRouters } from 'opentypebb'
+import { loadAllRouters } from '@traderalice/opentypebb'
 
 let _routeMap: Map<string, string> | null = null
 
diff --git a/src/plugins/mcp.ts b/src/server/mcp.ts
similarity index 100%
rename from src/plugins/mcp.ts
rename to src/server/mcp.ts
diff --git a/src/openbb/api-server.ts b/src/server/opentypebb.ts
similarity index 91%
rename from src/openbb/api-server.ts
rename to src/server/opentypebb.ts
index c05904b8..c4f28886 100644
--- a/src/openbb/api-server.ts
+++ b/src/server/opentypebb.ts
@@ -9,7 +9,7 @@
 import { Hono } from 'hono'
 import { cors } from 'hono/cors'
 import { serve } from '@hono/node-server'
-import { createExecutor, loadAllRouters } from 'opentypebb'
+import { createExecutor, loadAllRouters } from '@traderalice/opentypebb'
 
 export function startEmbeddedOpenBBServer(port: number): void {
   const executor = createExecutor()