Bitrix24 VibeCode: Building Business Apps with AI Agents

What Is Bitrix24 (Alaio) VibeCode and How Vibecoding Works

Bitrix24 VibeCode (vibecode.bitrix24.tech) is a four-layer platform - AI-optimised API, built-in AI models, isolated hosting, and ready-made AI agents - that eliminates the most expensive part of custom software: writing the code itself.

Vibecoding flips the traditional development workflow. Instead of a developer translating business requirements into code, you describe a task in plain language to an AI agent (Claude Code or Codex running on your desktop). The agent reads the Vibe API documentation, writes the application, and deploys it to a Black Hole server - all in one session.

The platform covers four layers that developers previously had to assemble by hand:

The three-step quick-start path:

1. Get a key (~2 min) - log in to vibecode.bitrix24.tech, authorise your Bitrix24 portal, and copy the API key.
2. Prompt the AI agent - open Claude Code or Codex, paste the starter prompt the platform provides (it includes your key and a pointer to GET /v1/me for self-documentation), and describe your task.
3. Deploy - tell the agent "deploy to server." It creates a Black Hole server, installs the app, and returns a live URL (https://app-xxxxxxxx.vibecode.bitrix24.tech/).

Step 2 is where all the work happens - and it happens outside the platform, in your AI tool. Claude Code and Codex need desktop/terminal access to write files and make requests; browser-only or Discord-based workflows won't work here.

VibeCode is not a visual no-code builder, and it does not include a coding AI by default - you bring your own model (Claude, Codex, or others via BYOK). The platform provides everything around that model: the API, the infrastructure, and the safety rails.

Vibe API: A Smarter REST Layer Built for AI

Vibe API wraps 300+ Bitrix24 methods into entity-level endpoints with automatic pagination, rate-limit queuing, MongoDB-style filters, and camelCase field mapping - removing the most common sources of AI coding errors against raw Bitrix24 REST.

Working directly with the Bitrix24 REST API in an AI-generated app creates predictable problems: the AI must handle pagination manually, respect rate limits (503/429 errors), map UPPER_CASE field names, and manage batch calls. Vibe API handles all of this automatically:

• Entity wrappers for 39 CRM entities, bots (31 methods), chats (16), workflows/business processes (5), telephony (10), triggers, timelines, warehouses, and the activity feed.
• Auto-pagination up to 5,000 records per call with date-window handling for large date ranges.
• Queue and batching: the platform assembles batches of 50+ operations, respects Bitrix24 rate limits internally - no 503/429 errors reach your app.
• MongoDB-style filters: { closedAt: { $gte: "2025-01-01", $lte: "2025-03-31" }, stageSemanticId: "F" } - readable by any AI model without extra prompt engineering.
• camelCase ↔ UPPER_CASE field mapping handled transparently.
• CRUD for custom fields (userfields) across deals, leads, contacts, companies, and quotes - no separate webhook setup needed.
• GET /v1/me - self-description endpoint: the key describes its own capabilities and available APIs so the AI model can "read the docs" and produce correct code on the first attempt.

The deliberate decision to remove a generic POST /v1/call endpoint in favour of entity wrappers means the AI always gets field transformation, pagination, and error handling for free.

For large-scale analytics (tens of entities, up to 10 million rows), the Bitrix24 BI Connector API is a simpler alternative - a dedicated BI API within VibeCode itself is on the roadmap.

An AI agent interacts with the VibeCode platform layers to build and deploy a Bitrix24 app.

flowchart LR
    DEV["Developer / Business Owner\n(plain-language task)"] --> AGENT["AI Agent\nClaude Code / Codex"]
    AGENT --> VAPI["Vibe API\n300+ Bitrix24 methods\nauto-pagination · batching\nMongoDB filters"]
    VAPI --> B24["Bitrix24 Portal\nCRM · Tasks · Chats\nWorkflows · Telephony"]
    AGENT --> ROUTER["AI Model Router\nBitrixGPT · BYOK\nOpenAI-compatible"]
    AGENT --> BH["Black Hole Server\nIsolated deploy\nSource escrow · Rollback"]
    BH --> B24
    MCP["MCP Server\n@vibe/mcp-vibe-api"] --> VAPI

Built-In AI Models and BYOK

VibeCode includes an OpenAI-compatible AI Model Router at a single endpoint - one key, one format - with several built-in models available at no extra cost, plus BYOK support for connecting your own OpenAI, Anthropic, or OpenRouter keys.

The router endpoint is https://vibecode.bitrix24.tech/v1/ai/chat/completions. Omitting the model parameter automatically routes to BitrixGPT 5.

Built-in models (selected)

BYOK (Bring Your Own Key): connect your existing OpenAI, Anthropic, or OpenRouter API keys. Traffic flows through your own account; the platform acts as a unified router with no additional charge for the routing itself. External models including Claude Opus 4, Claude Sonnet 4, Claude Haiku 4, and GPT-4o mini are reachable via the same single endpoint once BYOK is configured.

The Whisper model is particularly useful for transcribing voice messages and call recordings directly into Bitrix24 deal or contact records - see the AI call analysis capabilities in Bitrix24 for the broader context of how transcription feeds into CRM workflows.

Black Hole Servers: Secure, Isolated Deployment

Black Hole servers are the VibeCode hosting layer: each server is invisible from the public internet, communicates outward through an encrypted tunnel, and can be deployed in 2-3 minutes by instructing the AI agent - no SSH, no manual configuration.

Key characteristics:

• Network isolation: the server has no inbound public exposure. It initiates an outbound tunnel to reach Bitrix24 and to serve the app URL, so there is no open attack surface.
• Deploy by key: tell the agent "deploy to server" - it calls the VibeCode infra API, provisions the server, installs dependencies, and starts the app. The agent returns a live URL.
• Sleep / wake cycle: idle servers sleep automatically to reduce resource consumption and wake on the first incoming request (typically within a couple of minutes).
• Automatic source escrow: every deploy snapshots the source code. If a new version breaks something, a one-click rollback restores the previous working state.
• Galaxy mode: a Galaxy host packs roughly 25 Black Hole applications onto a single server, significantly reducing per-app infrastructure cost for teams running many small internal tools.
• Event subscriptions without polling: instead of polling event.get on a timer, the platform registers a handler on the Bitrix24 side and delivers each portal event (new task, deal change, etc.) directly to the Black Hole app through the tunnel. The platform will also wake a sleeping server to deliver an event.
• OPEN mode: an optional mode that adds direct SSH access for advanced debugging.
• S3-compatible storage at /v1/storage/* is available for app file storage.

For teams evaluating data sovereignty, it is worth noting that VibeCode also supports on-premise Bitrix24 deployments (see the on-premise section below) so the entire stack - portal, app logic, and AI - can run within your own infrastructure. You can find a detailed breakdown of self-hosted architecture options in Self-Hosted Bitrix24 on AWS, Azure, or a Private Cloud.

Hermes AI Agents and the CoALA Memory Model

Hermes is the current production AI agent in VibeCode - replacing the earlier OpenClaw prototype - and it features contextual memory based on the CoALA architecture, a knowledge graph, role-aware call transcription, and self-healing capabilities.

The agent design follows a "one personal agent per employee" model so that access keys and data contexts do not overlap between colleagues.

Key Hermes capabilities:

• CoALA contextual memory - the agent maintains a structured memory of previous interactions, not just a raw conversation log. This enables multi-session tasks without re-explaining context every time.
• Knowledge graph - entities encountered during work (deals, contacts, tasks) are linked, so the agent can answer cross-entity queries ("which contacts from this company have open deals in Stage 3?") without multiple explicit lookups.
• Self-learning - the agent refines its understanding of your portal's data structure over time.
• Role-aware transcription - call recordings are transcribed with speaker labels, not just raw text. Summaries identify who said what and what was agreed, suitable for writing into deal timeline entries.
• Self-healing - if the agent's server encounters an error, it attempts to recover automatically before alerting the user.

Agents are managed from within the Bitrix24 portal chat. A service key (valid for 3 days, scoped only to the agent's server, and revocable) is available for diagnostics when an agent needs repair.

The earlier OpenClaw prototype was retired after it proved unstable - users could trigger scenarios the development team had not anticipated. Hermes was designed with tighter guardrails and lower token consumption.

Cowork / Code: Your AI Assistant Inside Bitrix24

Cowork is a macOS/Windows desktop application pre-connected to your Bitrix24 portal under your own permissions - it functions as a Claude/ChatGPT-level AI assistant that already knows your portal data, with a separate Code tab for writing and deploying VibeCode apps.

Because the app authenticates with your portal credentials on launch, there is no separate setup step to access CRM data, chats, or tasks.

Cowork mode capabilities:

• Ask questions about your own portal data: "Analyse my last three months of chats and give me a communication style breakdown with per-contact statistics."
• Attach files (images, Word, Excel, PowerPoint, PDF) - the agent reads the content and answers questions about it.
• Call follow-up skill: after an internal call, Cowork generates an AI summary - what was discussed, what was agreed - and can write it to the relevant task or deal.
• Session folders with colour labels, keyboard shortcuts (⌘N new session, ⌘P/⌘F folder navigation), light/dark theme, and a live context-size indicator.

Code mode capabilities:

• Write, iterate, and deploy VibeCode apps directly from the desktop without switching to a terminal.
• View deployment status, the live app URL, deployment date, and project folders in a single panel.
• "Plan" panel for structuring multi-step development tasks.
• Deploy to Galaxy hosts without the 50% hang issue that affected earlier versions.

Cowork/Code is particularly valuable for Bitrix24 IT and software company teams that want AI-assisted development without leaving the Bitrix24 ecosystem.

Security Controls: Read-Only Keys and MCP

The most important safety mechanism in VibeCode is the read-only key mode: any write operation (create, update, delete) is blocked at the platform layer before it reaches Bitrix24, returning WRITE_BLOCKED_READONLY_KEY - protecting your CRM from accidental AI-driven data changes.

This matters because Bitrix24 does not log field-level CRM changes at the user-activity level by default. An AI model that silently overwrites deal fields may not be detected until a backup restore is the only option - and that loses recent data. Read-only keys eliminate this risk at the source.

Key types and their purposes

Recommended access policy:

• Set the default policy in /keys to "all new keys - read-only."
• Grant write access explicitly per employee or per application.
• The Companion bot sends a notification to the key owner whenever an admin changes key permissions (who changed it and when).

MCP server (@vibe/mcp-vibe-api): the Model Context Protocol server exposes 25 tools - portal management, key CRUD, entity CRUD (list/get/create/update/delete/search/batch/aggregate/fields), and discover (schema for 29 entities). With MCP configured, an AI model calls Vibe API as a structured tool rather than constructing raw HTTP requests, which reduces hallucination of endpoint names and parameters. Supports both stdio and HTTP modes (--http --port 3001).

Additional security layers:

• 2FA on portal login (prompted once per new device).
• Per-key request audit log.
• Store keys in environment variables, never in prompts or source code.

For teams with strict data governance requirements, the Self-Hosted Bitrix24 Security Hardening checklist covers complementary controls at the infrastructure level.

Practical Use Cases: What Teams Build with VibeCode

Based on documented patterns, the most common VibeCode use cases fall into three categories: CRM analytics dashboards (buildable in roughly 15 minutes), voice-to-task automation, and third-party data sync - all requiring no traditional developer.

CRM analytics dashboard (~15 minutes)

An AI agent is given a read-only key and instructed to build a dashboard showing deal counts, stage distribution, or activity-feed post volume grouped by employee and month. The agent calls POST /v1/deals/search with a date filter, aggregates the data, and deploys a web app to a Black Hole server. The result is a live URL embeddable in any browser or Bitrix24 placement.

Voice-to-task bot

A bot built on the VibeCode bot platform receives a voice message from an employee in Bitrix24 chat, sends the audio to the Whisper endpoint (/v1/audio/transcriptions), converts the transcript into a structured task, and creates it via POST /v1/tasks. The bot lives entirely within Bitrix24 - no external messenger needed. Unlike Telegram bots, the conversation and any attached files never leave the Bitrix24 perimeter.

Call transcription to deal card

Using the Whisper model or Bitrix24's existing call transcripts (where available), the agent writes a role-labelled transcript and an AI-generated summary to the deal timeline entry. This directly feeds the AI call analysis workflow in Bitrix24 and can trigger downstream automations (follow-up task, stage move, manager alert).

Bitrix24 ↔ Google Sheets sync

The agent uses Vibe API entity endpoints to pull deal or contact data and writes it to a Google Sheet via the Sheets API. MongoDB-style filters let the agent select only the relevant records (e.g., deals closed this quarter in a specific pipeline), avoiding the need for manual exports.

Customer base analysis

With a read-only key, an AI agent runs aggregations across contacts, companies, and deals to identify patterns - average deal cycle by segment, top-performing sources, inactive contacts - and presents results as a formatted report or embeds them in a dashboard.

These use cases align with broader Bitrix24 business process automation patterns, but VibeCode extends them into fully custom applications rather than pre-configured workflow templates.

VibeCode on Self-Hosted (On-Premise) Bitrix24

The Vibe Connect module extends VibeCode capabilities to self-hosted Bitrix24 installations, meaning teams that run Bitrix24 on their own servers can build AI-powered apps while keeping all data within their own infrastructure.

This is significant for organisations in the UAE, Brazil, and Portugal that operate under data residency requirements (UAE PDPL, Brazil's LGPD, Portugal/EU GDPR). The full vibecoding stack - Cowork/Code, the bot platform, AI agents, and app deployment - is available in the on-premise configuration.

Key points for on-premise deployments:

• The Vibe Connect module connects a self-hosted Bitrix24 instance to the VibeCode platform. As of publication, the module is in testing and approaching general availability.
• AI scenarios available in the cloud (CoPilot in chats, CRM text generation, smart task hints, call transcription, voice message transcription) carry over to the on-premise version.
• Data processed by the AI does not leave the company's infrastructure in the on-premise configuration - VibeCode apps run against your own Bitrix24 instance.
• For teams evaluating the self-hosted path, the Self-Hosted CRM data sovereignty guide and the GDPR-compliant CRM case for self-hosted Bitrix24 cover the broader compliance context.

Note: specific feature availability and module status change frequently during the platform's active development phase. Verify current status at https://vibecode.bitrix24.tech/docs before implementation planning.