llms.txt
# Powabase
> Postgres, RAG, and agents. One backend.
Powabase is the AI-native Supabase alternative: per-project Postgres, auth, and storage, with RAG that indexes documents on upload and agents built in.
> Full site text for ingestion: https://powabase.ai/llms-full.txt
## What it is
Powabase is the Postgres backend for AI apps. Every project gets its own database, auth, storage, and dedicated compute. Documents index on upload, agents call tools over HTTP or MCP, and all of it sits behind one REST API.
Each project gets a fully isolated stack: Postgres + pgvector, GoTrue auth,
object storage, realtime, and dedicated compute, plus an Agent Runtime, RAG
Pipeline, and Workflow engine. One REST API instead of a vector database, an
agent framework, and the glue code between them.
## Capabilities
### Backend (BaaS)
- Per-project Postgres with Row Level Security and pgvector for vector search.
- GoTrue authentication (email, OAuth, magic links, MFA).
- S3-compatible object storage with signed URLs and CDN.
- Realtime subscriptions over WebSockets (postgres_changes + broadcast channels).
- Direct PostgREST access to your public schema for app data.
### RAG (Retrieval)
- Document ingestion: PDFs, DOCX, images (with OCR), and URLs.
- VLM OCR scoring 91% on OlmOCR-Bench.
- Full RAG pipeline scoring 98.7% on FinanceBench.
- Four indexing strategies: ChunkEmbed, PageIndex (hierarchical tree),
GraphIndex (entity + cross-reference enrichment), Doc2JSON (structured
field extraction).
- Four retrieval methods: vector search (pgvector), full-text (BM25),
hybrid (RRF fusion), tree search (LLM reasoning over PageIndex).
- Cross-encoder reranking (Cohere, Jina, Voyage, others).
### Agents
- ReAct loop with multi-step tool calling and streaming SSE.
- Sessions persist multi-turn state, including reasoning traces and
retrieved context per turn.
- Automatic context compaction: when a session outgrows the model's context
window, older turns are summarized and recent turns stay verbatim. The full
transcript stays in Postgres.
- Sessions and runs are rows in the project's Postgres (ai.agent_sessions,
ai.agent_runs), queryable with SQL or /rest/v1 under Row Level Security.
- Hooks and approval flow for human-in-the-loop tool calls.
- Multi-agent orchestrations: a coordinator delegates to specialised
entity agents and synthesises their replies.
### Tools
- Built-in: db query/write, HTTP request, code execute, storage read/write,
web search, web scrape.
- Custom tools via HTTP endpoints with JSON Schema input definitions.
- External MCP servers attached to agents for runtime tool discovery.
### Workflows
- DAG-based automation with blocks: triggers, conditions, agent calls,
HTTP, code execution, platform_api, response.
- Visual builder + REST API. Deployed workflows expose an HTTP endpoint
with webhook, cron, and interval triggers.
- Natural-language Copilot generates and edits workflow graphs from
plain-English descriptions.
### LLM providers
OpenAI, Anthropic, Google, Mistral, and open-source models via OpenRouter.
Bring your own keys; keys are stored per-project, encrypted at rest. BYOK is
optional — you can also skip keys and pay for inference from your Powabase credits.
### Hosting & deployment options
Powabase offers two deployment models:
- **Powabase Cloud** — fully managed at app.powabase.ai.
- **Self-hosted (open source)** — Powabase is open source under Apache-2.0 at
https://github.com/powabase-ai/powabase. Run the full single-project stack on
your own infrastructure with one `docker compose up`. Enterprise adds a
commercially-licensed, supported edition with SLAs, a Kubernetes Helm chart,
and private managed hosting (https://powabase.ai/pricing#enterprise).
## Guides
- [What is backend as a service (BaaS)?](https://powabase.ai/backend-as-a-service/) — Backend as a service (BaaS) is a hosted platform that gives an app its server-side building blocks, such as a database, user authentication, file storage, realtime updates, and an API, so developers configure them instead of building and running them. Powabase is a BaaS for AI apps: per-project Postgres, auth, and storage, plus built-in RAG, agents, and workflows.
- [How do you self-host Supabase?](https://powabase.ai/self-hosted-supabase/) — Self-hosted Supabase is the open-source Supabase stack (Postgres, GoTrue auth, PostgREST, Realtime, Storage, Edge Functions, and Studio) running on your own servers, usually with Docker Compose, instead of on Supabase's managed platform. Powabase's Apache-2.0 self-host edition runs that same data plane unchanged and adds a RAG pipeline and agent runtime on top.
- [What is a vector database?](https://powabase.ai/vector-database/) — A vector database is a database that stores embeddings, the lists of numbers a machine learning model produces to represent the meaning of text, images, or audio, and finds the stored items closest to a query vector using approximate nearest-neighbor indexes. Powabase keeps vectors in pgvector inside each project's own Postgres, with retrieval built in.
- [How do you enforce access control in RAG with Row Level Security?](https://powabase.ai/rag-with-row-level-security/) — RAG with Row Level Security is a pattern where documents, chunks, and embeddings live in Postgres tables with owner or tenant columns, and RLS policies decide which rows each user's query can see, so vector search only returns passages the caller may read. We run every Powabase project on its own Postgres with RLS and pgvector, and knowledge-base search takes its scope on each request.
- [What does an AI agent need from a backend?](https://powabase.ai/ai-agent-backend/) — An AI agent backend is the server-side layer that runs an agent's reason-and-act loop and supplies what the model alone lacks: tools it can call, sessions that remember earlier turns, context management for long conversations, approval gates for risky actions, streaming to the user, and a record of every run. Powabase ships that runtime built in, next to each project's Postgres.
- [What is AI workflow automation?](https://powabase.ai/workflow-automation/) — AI workflow automation is running a business process as a defined sequence of steps, started by a trigger such as a webhook, a schedule, or an API call, where some steps call AI agents or models and the rest fetch data, branch on conditions, run code, and call other systems. Powabase runs these workflows in the same project that holds your data, documents, and agents.
- [How do you migrate from Supabase to Powabase?](https://powabase.ai/migrate-from-supabase/) — Migrating from Supabase to Powabase means dumping your own Postgres schemas and data, restoring them into a Powabase project in one transaction, pointing your app at the new URL and keys, and moving auth users, stored files, Edge Functions, and pg_cron jobs as their own steps. Powabase runs the same PostgREST, GoTrue, Storage, and Realtime services, so most app code carries over unchanged.
## Comparisons
- [Supabase alternative](https://powabase.ai/supabase-alternative/) — Powabase runs the same open-source core services as Supabase (GoTrue, PostgREST, Storage, and Realtime), and both give every project its own Postgres instance. We add managed document extraction, retrieval, an agent runtime, and workflows on top, so your AI features live next to your data. Building RAG or agents? Pick Powabase. If a database with auth and storage is all you need, Supabase is enough.
- [Firebase alternative](https://powabase.ai/firebase-alternative/) — Firebase is Google's app platform. It centers on Firestore, a NoSQL database, and has strong mobile SDKs. Powabase is a Postgres backend with document ingestion, retrieval, and an agent runtime built in, and every project gets its own Postgres instance. If your AI app needs SQL plus RAG and agents in one place, pick Powabase. Building mobile-first around Firestore and Google's SDKs? Firebase fits.
- [Convex alternative](https://powabase.ai/convex-alternative/) — Convex is a reactive backend. You write queries, mutations, and actions in TypeScript against its own document-relational database. Powabase gives every project its own Postgres instance with SQL, auth, storage, and realtime, and we manage document ingestion, retrieval, and agents on top. Pick Powabase if your AI app needs SQL and managed RAG. For a TypeScript-first reactive backend, Convex is built for it.
- [Neon alternative](https://powabase.ai/neon-alternative/) — Neon is serverless Postgres with branching and scale-to-zero, plus backend services such as auth, a PostgREST-compatible Data API, and Functions. Powabase gives every project its own Postgres instance with auth, REST, storage, and realtime, then adds managed document ingestion, retrieval, agents, and workflows. Building RAG or agents? Pick Powabase. If branching and scale-to-zero matter most, pick Neon.
- [Pinecone alternative](https://powabase.ai/pinecone-alternative/) — Pinecone is a managed vector database with serverless indexes, hosted embedding models, and Pinecone Assistant for chat over files. Powabase keeps your vectors in pgvector inside each project's own Postgres, right next to your app data, and adds document extraction, auth, storage, and agents. Want one backend for your whole RAG app? Pick Powabase. If a vector index is all you need, pick Pinecone.
- [LangChain alternative](https://powabase.ai/langchain-alternative/) — LangChain is an open-source framework for building agents in your own code. A production RAG app built on it usually needs a database and auth too, such as Supabase, plus a vector store such as Pinecone. Powabase puts Postgres, auth, storage, document ingestion, retrieval, and a managed agent runtime behind one API. Pick Powabase to skip the assembly, or LangChain for full control in code.
## Pricing
Powabase uses usage-based pricing, not per-seat. Each monthly plan converts its
fee into a pool of credits you spend on actual usage — compute hours, per-call
actions, active users, and storage — and unused credits roll over month to month.
Paid plans add prepaid credits and progressively cheaper unit rates; underneath
every plan is the same pay-as-you-go meter.
Plans:
- **Free** — $0. All platform features. Pay-as-you-go on usage with no
monthly minimum and no commitment. $10 in free credits on sign-up.
Free projects run on Sandbox compute only.
- **Self Serve** — $25 / month. $25 in monthly credits to spend on
compute hours, per-call costs, MAU overages, and storage. Pick any compute
tier (Sandbox through Foundry). Up to 25% cheaper per-call costs and 15%
cheaper per-hour compute vs. Free. Email support.
- **Scale** — $300 / month. $300 in monthly credits. Up to 50% cheaper
per-call costs and 20% cheaper per-hour compute vs. Free. Lowest
overage rates. Priority live support.
- **Enterprise** — Custom. Managed on Powabase Cloud or private hosting
in your AWS / GCP / Azure / on-prem environment. $500 in platform
credits to start. SOC 2, DPA, MSA, regional residency, SSO, RBAC,
audit logs, BYO KMS / IdP, air-gapped supported, dedicated solutions
engineer.
See https://powabase.ai/pricing for the live pricing matrix.
## Project isolation
Each Powabase project is a dedicated stack with its own resources and
backups. There is no shared vector DB, no noisy neighbours, and no
cross-tenant traffic. The control plane manages routing and authentication;
the data plane is isolated per project.
## Powabase MCP server
The Powabase MCP server is live at https://mcp.powabase.ai/mcp. Any MCP client
that supports remote servers — Claude Code, Cursor, and others — connects over
streamable HTTP and signs in with OAuth. In Claude Code:
claude mcp add --transport http powabase https://mcp.powabase.ai/mcp
Connected agents can list and inspect projects, run SQL, manage auth users and
storage, and create and run agents, knowledge bases, orchestrations, and
workflows. Append `?read_only=true` to the URL to expose only tools that
don't write.
## Integrations — coding agents & vibe-coding platforms
Powabase connects to the AI coding tools you already use. Coding agents
(Claude Code, Codex, Antigravity, OpenCode, Replit) install the Powabase
Agent Skill — `npx skills add powabase-ai/agent-skills` — and build a full
backend from a prompt. Browser-based vibe-coding platforms (Replit, Lovable,
v0, Base44, Bolt.new) connect over the REST API using the Connect dialog's
Project URL + Service Role (Secret) Key (server-side only; the Anon key is the
client-side, RLS-respecting key). A third option is the auth-connection guide
at https://docs.powabase.ai/guides/auth-connection. Coding agents can also connect
through the Powabase MCP server described above.
Per-tool connector pages with quick starts and example prompts live at
https://powabase.ai/integrations.
## Blog
Writing from the Powabase team, by topic (full text at https://powabase.ai/llms-full.txt):
### RAG & Retrieval — https://powabase.ai/blog/category/rag/
Indexing, retrieval, and knowledge bases on Postgres.
- [How to Make Multitenant Vector Indices Scalable](https://powabase.ai/blog/how-to-make-multitenant-vector-indices-scalable/) — 2026-10-08. Multitenant vector indices can bottleneck fast-growing apps. Learn proven strategies to scale them efficiently without sacrificing speed or isolation.
- [Multi-Tenant RAG Tenant Isolation](https://powabase.ai/blog/multi-tenant-rag-tenant-isolation/) — 2026-09-02. Master multi-tenant RAG tenant isolation with proven strategies to keep data secure, prevent leakage, and scale confidently across all your customers.
- [Pinecone Alternative: Migrate to pgvector in an Afternoon](https://powabase.ai/blog/pinecone-alternative-migrate-to-pgvector-in-an-afternoon/) — 2026-08-13. Pinecone's $50/month minimum killed hobby RAG. This Pinecone alternative walks you through migrating to pgvector with vec2pg, schema, code, and cost math.
- [RAG Backend, Not Framework: Agentic Loops on Postgres](https://powabase.ai/blog/rag-backend-not-framework-agentic-loops-on-postgres/) — 2026-08-13. An agentic RAG loop's real dependency is a governed RAG backend: pgvector, BM25, metadata, and state on one Postgres, not LangGraph glue.
- [LangChain Alternative in 2026: The Production RAG Stack](https://powabase.ai/blog/langchain-alternative-in-2026-the-production-rag-stack/) — 2026-07-08. A production RAG audit says you don't need LangChain in 2026. The alternative: native SDK, Postgres/pgvector, and a thin router you don't build.
### Agents & Workflows — https://powabase.ai/blog/category/agents/
Building, running, and automating agents with a database behind them.
- [Agentic AI in Wearables: How Autonomous Agents Work](https://powabase.ai/blog/agentic-ai-in-wearables-how-autonomous-agents-work/) — 2026-10-15. Agentic AI in wearables explained: multi-agent frameworks, on-device vs. cloud inference, personal health coaching, smart glasses, privacy, and what's next.
- [The State of Embodied AI in 2026: A Complete Guide](https://powabase.ai/blog/the-state-of-embodied-ai-in-2026-a-complete-guide/) — 2026-10-13. Explore the state of embodied AI in 2026: how VLA models, world models, and LLMs power robots, plus market size, challenges, and the road to AGI.
- [Agentic AI Trading Strategy: A Step-by-Step Guide](https://powabase.ai/blog/agentic-ai-trading-strategy-a-step-by-step-guide/) — 2026-10-09. Master an agentic AI trading strategy with our clear, step-by-step guide. Learn how to build, test, and deploy autonomous trading systems that act on your
- [How to Build a Hedge Fund Trading Model with Agentic AI](https://powabase.ai/blog/how-to-build-a-hedge-fund-trading-model-with-agentic-ai/) — 2026-10-07. Learn how to build a hedge fund trading model with agentic AI: design multi-agent roles, wire data APIs, manage risk, backtest, and deploy to paper trading.
- [Prompt Caching for AI Agents: Where the Savings Come From](https://powabase.ai/blog/prompt-caching-for-ai-agents-where-the-savings-come-from/) — 2026-10-02. Learn how prompt caching for AI agents cuts cost and latency in long loops — KV cache basics, provider differences, cache hit rates, and best practices.
- [Best Automation Tool 2026: viaSocket vs Zapier, Make, n8n & Powabase](https://powabase.ai/blog/best-automation-tool-2026-viasocket-vs-zapier-make-n8n-powabase/) — 2026-09-10. Searching for the best automation tool 2026? We compare viaSocket, Zapier, Make, n8n, and Powabase so you can pick the right fit for your workflows.
- [Agent Backend: One Postgres vs. Redis + Kafka + Pinecone](https://powabase.ai/blog/agent-backend-one-postgres-vs-redis-kafka-pinecone/) — 2026-09-03. A 200-line Postgres agent orchestrator shows your database can be the framework. What that agent backend gets right, and what production still needs.
- [Build an AI Customer Support Agent on Postgres](https://powabase.ai/blog/build-an-ai-customer-support-agent-on-postgres/) — 2026-08-27. Build an AI customer support agent on Postgres: tool-calling, pgvector semantic search, escalate-on-no-match, and one database instead of Pinecone plus a CRM.
- [Long-Running Agents on Postgres: Durable, Resumable State](https://powabase.ai/blog/long-running-agents-on-postgres-durable-resumable-state/) — 2026-08-27. How to build long-running agents on Postgres with durable, resumable state, so workflows survive failures and pick up where they left off.
- [Agent Memory in Postgres: One DB, No Vector Store](https://powabase.ai/blog/agent-memory-in-postgres-one-db-no-vector-store/) — 2026-08-19. Store agent memory in Postgres without a separate vector store. Learn how one database handles everything your AI agent needs to remember.
### Backend & Comparisons — https://powabase.ai/blog/category/backend/
Backend-as-a-service for AI apps, and how the options compare.
- [Row-Level Security for AI Agents: Run as User or Not?](https://powabase.ai/blog/row-level-security-for-ai-agents-run-as-user-or-not/) — 2026-10-02. Row-level security for AI agents raises a key question: should they run as the user or not? Explore the tradeoffs to make the right call for your system.
- [Claude Workspace vs Custom Build: When to Build](https://powabase.ai/blog/claude-workspace-vs-custom-build-when-to-build/) — 2026-09-24. Claude workspace vs custom build: learn where Claude Projects are enough, where they hit a wall, and when to build custom with Powabase.
- [Vibe Coding Backend Risks: Why Building From Scratch Fails](https://powabase.ai/blog/vibe-coding-backend-risks-why-building-from-scratch-fails/) — 2026-09-24. Vibe coding backend risks like security holes and tenant data leaks make from-scratch builds fragile. See why Powabase makes core backend logic reliable.
- [Best Backend as a Service 2026: Firebase vs Supabase & More](https://powabase.ai/blog/best-backend-as-a-service-2026-firebase-vs-supabase-more/) — 2026-09-17. The best backend as a service compared: Firebase, Supabase, Neon, Appwrite, and Powabase on pricing, features, AI, and lock-in, plus which to pick.
- [AI App Development Agency Backend: Own the Recurring Layer](https://powabase.ai/blog/ai-app-development-agency-backend-own-the-recurring-layer/) — 2026-07-30. Why AI app agencies are moving from one-off builds to managed hosting and MRR: own a governed multi-tenant backend instead of reselling SaaS.
- [CI/CD for AI: Building Pipelines for ML and LLM Workflows](https://powabase.ai/blog/ci-cd-for-ai-building-pipelines-for-ml-and-llm-workflows/) — 2026-07-30. CI/CD for AI demands more than standard DevOps. Learn how to build robust pipelines that handle the unique challenges of ML and LLM workflows.
- [Backend as a Service: Beat the AI Cleanup Economy](https://powabase.ai/blog/backend-as-a-service-beat-the-ai-cleanup-economy/) — 2026-07-23. Agencies pay $200–400/hr to fix unreviewed AI code. How an agent-native backend as a service ships correct schemas, RLS, and migrations.
- [Why Vibe Coding Platforms Need a Ro
