RESTai

AIaaS (AI as a Service) — Create AI projects and consume them via a simple REST API.

---

Live Demo

Try RESTai without installing — ai.restai.cloud

Login: demo / demodemo (restricted account — can browse and chat, but cannot create or modify projects)

---

Quick Start

Install from PyPI

pip install restai-core
restai init      # Create database + admin user
restai migrate   # Run migrations
restai serve     # → http://localhost:9000/admin (admin / admin)

Use an env file for configuration:

restai serve -e .env -p 8080 -w 4

Available on PyPI — includes the pre-built React frontend, no Node.js required.

Run from source (development)

git clone https://github.com/apocas/restai && cd restai
make install
make dev  # → http://localhost:9000/admin (admin / admin)

Docker

Pull the official prebuilt image (multi-arch — linux/amd64 and linux/arm64):

docker run -p 9000:9000 apocas/restai:latest
# → http://localhost:9000/admin (admin / admin)

Also published to GitHub Container Registry as ghcr.io/apocas/restai:latest. Pin a version with :6.2.13 (or :6.2, :6) instead of :latest.

Or build locally with the bundled compose stack:

docker compose --env-file .env up --build

Updating

PyPI:

pip install --upgrade restai-core
restai migrate -e .env

From source:

make update

Fetches the latest release tag from GitHub, installs dependencies, runs database migrations, and rebuilds the frontend. Auto-detects GPU for GPU-specific deps.

---

Why RESTai?

• Multi-project AI platform — RAG (with optional SQL-to-NL and auto-sync from URLs/S3), Agents, Block (visual logic), and Inference in one place
• Full Web UI included — React dashboard with analytics, not just an API
• Any LLM — OpenAI, Anthropic, Ollama, Gemini, LiteLLM, vLLM, Azure, and more
• Feature complete — Teams, RBAC, OAuth/LDAP, TOTP 2FA, token tracking, per-project rate limiting, white-label branding, Kubernetes-native
• Extensible tools — MCP (Model Context Protocol) for unlimited agent integrations
• Token tracking, cost & latency analytics — built-in dashboard with daily usage, per-project costs, latency monitoring, and top LLM charts

---

Features

Dashboard & Analytics

Track token usage, costs, latency, and project activity from a centralized dashboard. Daily charts for tokens, costs, and response latency per project — identify performance regressions at a glance.

Projects & Chat Playground

Create and manage AI projects. Each project has its own LLM, system prompt, tools, and configuration. Test instantly in the built-in chat playground.

RAG (Retrieval-Augmented Generation)

Upload documents and query them with LLM-powered retrieval. Supports multiple vector stores, reranking (ColBERT / LLM-based), sandboxed mode to reduce hallucination, and evaluation via deepeval. Optionally connect a MySQL or PostgreSQL database to translate natural language questions into SQL queries automatically.

Knowledge Graph (RAG)

Opt-in entity extraction layered on top of RAG. When enabled on a project, every ingested document is run through a NER pipeline (dslim/bert-base-NER by default) and the extracted people, organizations, locations, and other entities are persisted in a queryable graph alongside the vector store. Entity extraction runs as a background task so ingestion stays fast.

What you get:

• Entity-aware retrieval — A custom postprocessor boosts retrieved chunks whose source documents mention entities found in the user's query, improving recall on questions that reference specific names
• Visual graph explorer — Interactive force-directed graph (vis-network) of all entities and their co-occurrence relationships. Click any node for source list, mention count, and related entities
• Entity browser — Searchable, type-filtered table of every entity in the project with mention counts and one-click rename / merge / delete
• Disambiguation — "Find Duplicates" surfaces near-identical entity pairs (Levenshtein-based) for one-click merging — fix "Acme", "Acme Corp", "ACME Inc." in seconds
• Natural language graph queries — Ask questions like "What did the document say about Acme Corp?" and get answers scoped to entity-matched sources, bypassing pure vector similarity
• Rebuild — Re-extract the entire graph from existing sources after enabling the feature on a populated project

Enable it from the project edit page (Knowledge tab → "Enable Knowledge Graph"). Available exclusively for RAG projects.

Agents + MCP

Zero-shot ReAct agents with built-in tools and MCP (Model Context Protocol) server support for extensible tool access. Connect any MCP-compatible server via HTTP/SSE or stdio.

Agentic Browser

Give your agents a real headless Chromium they can drive — log in to vendor portals, fill forms, scrape data, download invoices, take screenshots. Powered by Playwright running in a per-chat Docker container with cookie / localStorage persistence so the agent only needs to log in once.

Nine browser_* builtin tools the LLM composes into workflows:

• browser_goto, browser_click, browser_fill, browser_select, browser_wait
• browser_content (sanitized HTML/markdown of the current page)
• browser_screenshot (rendered inline in the chat — uses the same image cache as draw_image)
• browser_download (files land in the container's /home/user/downloads/ so the terminal tool can pick them up)
• browser_eval (admin-opt-in JS escape hatch)

Built for production, safely:

• 🔐 Encrypted secrets vault — admin stores credentials per-project (portal_password, etc.) and the agent calls browser_fill(selector, secret_ref="portal_password"). The plaintext is resolved server-side and typed straight into the browser; it never enters the LLM's context, the inference log, the audit log, or the chat transcript.
• 🛡️ Per-project domain allowlist — browser_goto refuses anything not on browser_allowed_domains (supports *.example.com suffix globs). Defends against prompt injection that tells the agent to navigate to a hostile site.
• 💾 Persistent sessions — login state (cookies + localStorage) is saved to Redis keyed by (project_id, domain) with a 30-day TTL. Future chats on the same project skip the login dance.
• 🧹 Auto-cleanup — idle browser containers are reaped by crons/browser_cleanup.py after browser_timeout seconds (default 15 min).

Toggle the feature on at Settings → Agentic Browser. Reuses the same Docker daemon as the sandboxed terminal; image defaults to mcr.microsoft.com/playwright/python:v1.48.0-jammy.

Inference (Multimodal)

Direct LLM chat and completion. Supports sending images alongside text using any vision-capable model (LLaVA, Gemini, GPT-4o, etc.).

Block (Visual Logic Builder)

Build processing logic visually using a Blockly-based IDE — no LLM required. Drag-and-drop blocks to define how input is transformed into output. Use the "Call Project" block to invoke other RESTai projects, enabling composition of AI pipelines without writing code.

Supported blocks: text operations, math, logic, variables, loops, and custom RESTai blocks (Get Input, Set Output, Call Project, Log).

Classifier Playground

Built-in zero-shot text classifier with multiple model support. Enter any text and a comma-separated list of candidate labels — get instant classification scores without training. Use it standalone from the UI or programmatically via POST /tools/classifier. Also available as a Blockly block for visual logic projects.

Available models:

• BART Large MNLI (default) — Established standard, reliable general-purpose classifier
• DeBERTa v3 Large — Best accuracy, newer architecture
• DeBERTa v3 Base — Good balance of speed and accuracy
• XLM-RoBERTa — Multilingual support for 100+ languages
• DistilBERT MNLI — Fastest and lightest, ideal for high-volume classification

MCP Server

RESTai includes an optional built-in MCP (Model Context Protocol) server that exposes your projects as tools consumable by any MCP client — Claude Desktop, Cursor, or custom agents. Each user authenticates with a Bearer API key and can only access their assigned projects.

Enable via MCP_SERVER=true environment variable or the admin settings page (requires restart). Clients connect to http://your-host:9000/mcp/sse.

Available tools:

• list_projects — Discover which AI projects you have access to
• query_project — Send a question (with optional image) to any accessible project

Evaluation Framework

Built-in evaluation system to measure and track AI project quality over time. Create test datasets with question/expected-answer pairs, run evaluations with multiple metrics, and visualize score trends.

Metrics (powered by DeepEval):

• Answer Relevancy — Is the answer relevant to the question?
• Faithfulness — Is the answer grounded in the retrieved context? (RAG projects)
• Correctness — Does the answer match the expected output?

Prompt Versioning

Every system prompt change is automatically versioned. Browse the full history, compare versions, and restore any previous prompt with one click. Eval runs are linked to prompt versions, enabling A/B comparison — see exactly how a prompt change affected quality scores.

Image Generation

Local and remote image generators loaded dynamically. Supports Stable Diffusion, Flux, DALL-E, RMBG2, and more. Atlas Cloud and MuAPI are available as first-class asynchronous providers: select the provider in the Image Generators admin page, enter its API key, and choose a supported image model. MuAPI (documentation) accepts the live catalog's model or endpoint path and defaults to flux-dev.

GPU Auto-Detection & Management

RESTai automatically detects NVIDIA GPUs at startup and displays detailed hardware information in the admin settings — model name, VRAM, temperature, utilization, power draw, driver and CUDA versions. GPU support is auto-enabled when hardware is detected, or can be toggled manually.

make install also detects GPUs automatically and installs GPU dependencies when available.

Direct Access (OpenAI-Compatible)

Use LLMs, image generators, and audio transcription directly via OpenAI-compatible API endpoints — no project required. Team-level permissions control which models each user can access, and all usage counts toward team budgets.

Supported endpoints:

• POST /v1/chat/completions — Chat with any LLM (streaming supported)
• POST /v1/images/generations — Generate images via DALL-E, Flux, Stable Diffusion, etc.
• POST /v1/audio/transcriptions — Transcribe audio files

Works with any OpenAI-compatible SDK:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:9000/v1", api_key="YOUR_API_KEY")
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
)

Teams & Multi-tenancy

Each team has its own users, admins, projects, and LLM/embedding access controls — including image and audio generator permissions. Users can belong to multiple teams, each with optional custom branding.

Custom Branding (White-Labeling)

Each team can customize the platform appearance for its members — ideal for white-labeling or multi-tenant deployments where different teams need distinct identities.

Configurable per team:

• App Name — Override the platform name in the sidebar and header
• Logo — Custom logo via URL or data URI (replaces the default logo)
• Primary & Secondary Colors — Full MUI theme color override with live color picker
• Welcome Message — Custom landing text for team members

Multi-team users: When a user belongs to multiple branded teams, a team switcher appears in the sidebar letting them choose which branding to apply. The preference is persisted in user settings.

API: GET /teams/{id}/branding returns team branding without authentication (useful for custom login pages). Update branding via PATCH /teams/{id} with a branding object.

Two-Factor Authentication (TOTP)

Secure local user accounts with TOTP-based two-factor authentication, compatible with Google Authenticator, Authy, and other authenticator apps.

• User self-service — Enable/disable 2FA from the user profile page with QR code setup and one-time recovery codes
• Admin enforcement — Platform admins can enforce 2FA for all local users via the settings page (users cannot disable when enforced)
• Recovery codes — 8 single-use codes generated during setup for account recovery if the authenticator is lost
• Local auth only — 2FA applies to username/password login; SSO and API key authentication are unaffected
• Encrypted secrets — TOTP secrets are encrypted at rest using Fernet; recovery codes are stored as SHA-256 hashes

Guardrails

Protect your AI projects with input and output guards. Guards are regular RESTai projects — define safety rules via system prompts, and they'll evaluate every request and response automatically.

• Input Guard — Checks user questions before inference
• Output Guard — Checks LLM responses after inference
• Block or Warn mode — Hard-block unsafe content or flag it while passing through
• Analytics dashboard — Track block rates, view blocked requests, and monitor guard effectiveness over time

Audit Log

Every mutation (create, update, delete) across the platform is automatically logged — who did what, when, and which resource was affected. Admins can review the full audit trail from the admin dashboard.

Rate Limiting

Set per-project request limits to prevent abuse and control costs. Configure the maximum number of requests per minute in the project edit page. Returns HTTP 429 when the limit is exceeded.

Knowledge Base Sync

Automatically keep your RAG knowledge base up-to-date by syncing from external sources on a schedule. Configure per project — each project manages its own sources and sync interval.

Supported sources:

• Web URLs — Periodically re-scrape web pages and update the knowledge base
• Amazon S3 — Sync documents from S3 buckets with optional prefix filtering (supports all file types: PDF, DOCX, CSV, etc.)
• Confluence — Sync all pages from a Confluence Cloud space (HTML stripped to plain text, paginated)
• SharePoint / Microsoft 365 — Sync files from a SharePoint Online document library via Microsoft Graph API (OAuth2 client credentials, optional folder filter)
• Google Drive — Sync files from a Drive folder via service account (supports native Google Docs/Sheets/Slides export + binary files)

Features:

• Configurable sync intervals (15 minutes to 24 hours)
• Multiple sources per project with independent settings (splitter, chunk size)
• Manual "Sync Now" trigger via UI or API (POST /projects/{id}/sync/trigger)
• Last sync timestamp tracking
• All credentials masked in API responses (S3 keys, Confluence tokens, SharePoint secrets, Drive service account JSON)

Embeddable Chat Widget

Add an AI chat bubble to any website with a single <script> tag — no frontend development needed. The widget connects to a RESTai project, streams responses in real-time, and maintains conversation context.

<script
  src="https://your-restai.com/widget/chat.js"
  data-project-id="7"
  data-api-key="sk-abc123..."
  data-title="Support Bot"
  data-primary-color="#6366f1"
  data-welcome-message="Hi! How can I help?"
></script>

Customizable: title, subtitle, colors, avatar, position (left/right), and welcome message — all via data-* attributes. Configure and preview live from the Widget tab in the project page before deploying.

Secure: Requires a read-only, project-scoped API key (visible in page source by design — same model as Stripe publishable keys). Shadow DOM isolates styles from the host page.

WordPress Plugin

A full-featured WordPress plugin that turns any RESTai instance into the AI engine of a WordPress site. Each capability maps to its own RESTai project, so models, prompts and budgets stay tunable per task — and the plugin auto-provisions the starter projects on first connect, so there's nothing to wire up by hand.

What it does:

• Generate post content, excerpts and SEO meta straight from Gutenberg — also wires title/description/focus keyphrase into Yoast and Rank Math fields
• One-click featured image generation using whichever generator your team has access to (Flux, SDXL, DALL·E…)
• Translate any post to N languages as drafts, with Polylang/WPML compatibility
• AI comment moderation — auto-flag spam and toxic comments before they're approved, with optional suggested replies
• WooCommerce product descriptions and FAQ generation from product attributes
• Knowledge sync — every published post and page is auto-pushed into a Support Bot RAG project so the bot is always current
• AI site search — replaces native WP search with semantic answers from the support bot
• Embeddable chat widget — one toggle adds the chat bubble to public pages with the right widget key auto-provisioned
• AI-personalised transactional emails via the wp_mail filter
• Token usage and cost panel mirrored straight into the WP admin

One Make-style install: drop the plugin zip into Plugins → Add New → Upload Plugin, paste your RESTai URL + API key in Settings → RESTai, pick a team, click Auto-provision starter projects — done.

Telegram, Slack & WhatsApp Integration

Connect any project to Telegram, Slack, or WhatsApp — messages are processed through the project's chat pipeline and responses are sent back automatically.

• Telegram — Paste your BotFather token in the project settings. No public URL required.
• Slack — Create a Slack app at api.slack.com, add the standard message/history scopes, install it to your workspace, and paste the Bot Token (xoxb-...) in the project settings. Polled by the cron runner — no public URL or daemon required.
• WhatsApp Business Cloud API — First-party Meta integration (no Twilio middleman, ~1k free conversations/month). Webhook-driven: paste the auto-generated webhook URL into Meta Business Suite → WhatsApp → Configuration → Webhooks, configure phone-number id + access token + app secret + verify token in the project's Integrations tab, and the bot is live. Per-project allowlist of E.164 senders protects your number's quality rating from spam. Includes a send_whatsapp built-in tool for outbound notifications (constrained by Meta's 24-hour customer-service window). Requires a public URL — local dev needs a tunnel like ngrok or Cloudflare Tunnel.

Project Routines

Schedule recurring messages that auto-fire on any project — the message runs through the project's normal chat/question pipeline, so it works with RAG, agents, and block projects alike.

Settings & Configuration

White-label the UI, configure currency for cost tracking, set agent iteration limits, manage LLM proxy, and more.

---

Supported LLMs

Talks directly to provider SDKs. Each model has a configurable context window with automatic chat memory management — older messages are summarized rather than dropped.

---

Tech Stack

Backend: FastAPI · SQLAlchemy · LlamaIndex · Alembic Frontend: React 18 · MUI v5 · Redux Toolkit Vector Stores: ChromaDB · PGVector · Weaviate · Pinecone Databases: SQLite (default) · PostgreSQL · MySQL Package Manager: uv

---

API

All endpoints are documented via Swagger.

Create a project:

curl -X POST http://localhost:9000/projects \
  -u admin:admin \
  -H 'Content-Type: application/json' \
  -d '{
    "name": "my-rag",
    "type": "rag",
    "llm": "gpt-4o",
    "embeddings": "text-embedding-3-small",
    "vectorstore": "chroma"
  }'

Chat with a project:

curl -X POST http://localhost:9000/projects/1/chat \
  -u admin:admin \
  -H 'Content-Type: application/json' \
  -d '{"message": "What is RESTai?"}'

---

Installation

RESTai uses uv for dependency management. Python 3.11+ required.

Local

make install    # Install deps, initialize DB, build frontend
make dev        # Development server with hot reload (port 9000)
make start      # Production server (4 workers, port 9000)

Default credentials: admin / admin (configurable via RESTAI_DEFAULT_PASSWORD).

Docker

Prebuilt image (recommended — no build, no toolchain):

docker run -d --name restai -p 9000:9000 apocas/restai:latest

Pass an env file to inject configuration (API keys, DB host, etc.) and a volume to persist the SQLite DB / uploads across restarts:

docker run -d --name restai -p 9000:9000 --env-file .env \
  -v restai-data:/app/data \
  apocas/restai:6.2.13

Published on every release to both registries — pick whichever you prefer:

Built for linux/amd64 and linux/arm64 (Apple Silicon / Graviton). Pin a specific version in production (:6.2.13) rather than :latest. The publishing workflow lives in .github/workflows/docker-publish.yml — releases auto-build, or trigger manually with a tag input.

Build locally with compose (for development):

# Edit .env with your configuration, then:
docker compose --env-file .env up --build

Optional profiles for additional services:

docker compose --env-file .env --profile redis up --build      # + Redis
docker compose --env-file .env --profile postgres up --build   # + PostgreSQL
docker compose --env-file .env --profile mysql up --build      # + MySQL

Kubernetes (Helm)

A Helm chart is provided in chart/restai/.

helm install restai chart/restai/ \
  --set config.database.postgres.host=my-postgres \
  --set secrets.postgresPassword=mypassword

For production with multiple replicas, set fixed secrets for JWT and encryption:

helm install restai chart/restai/ \
  --set config.database.postgres.host=postgres \
  --set secrets.postgresPassword=mypassword \
  --set secrets.authSecret=$(openssl rand -base64 48) \
  --set secrets.ssoSecretKey=$(openssl rand -base64 48) \
  --set secrets.fernetKey=$(python -c 'from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())')

See chart/restai/ for full Helm values and configuration options.

Bulk ingest needs a shared volume. The API pod stages each upload to disk and the cron pod reads it back to index it, so the two must see the same filesystem. Set uploadStaging.enabled=true (a ReadWriteMany PVC, or a single-node RWO with both pods pinned to that node) — otherwise every bulk-ingest job stays queued with a file path the cron can't open:

helm install restai chart/restai/ \
  --set uploadStaging.enabled=true \
  --set uploadStaging.size=50Gi \
  --set uploadStaging.storageClass=nfs-client

Uploads behind a reverse proxy

The upload ceiling is an admin setting (Admin → Settings → Limits → Max Upload Size, default 100 MB), but a proxy in front of RESTai enforces its own limit first and returns an HTML 413 the app never sees. Raise both together:

Two more things scale with the ceiling. A single large upload occupies one worker for its whole duration, so run more than one (make start uses 4; the Docker entrypoint runs a single worker) and set a proxy read timeout that matches how long a big upload actually takes on your link. And staging is disk-backed: RESTai refuses a batch that would leave under 512 MB free (HTTP 507) rather than filling the volume, so size uploadStaging for the largest batch you expect to have in flight at once.

---

Architecture

Stateless (Production)

No state stored in the RESTai service — ideal for horizontal scaling.

Stateful (Development)

Direct interaction with the GPU layer — ideal for small deployments.

---

Environment Variables

LLM credentials (OpenAI, Anthropic, Gemini, Grok, Azure, etc.) live on each LLM in /admin/llms — encrypted per-LLM in the DB and scoped to teams. Image-generator credentials (DALL-E, Imagen) live in /admin/image-generators. GPU toggle, Redis (chat memory) and the vector backends (ChromaDB / PGVector / Weaviate / Pinecone) are configured in /admin/settings. None of these need env vars; GPU is auto-detected on first boot.

Full configuration in restai/config.py.

---

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

make dev        # Run dev server
pytest tests    # Run tests
make code       # Format with black

Note: This project started as 100% human-written code. Nowadays, only a small percentage of the codebase is human-developed — the majority is AI-generated.

---

License

Pedro Dias - @pedromdias

Licensed under the Apache License, Version 2.0. See LICENSE for details.