AgentSwarms: self-hosted agentic AI platform launches open source

AgentSwarms is a complete, self-hostable platform for agentic AI and business intelligence, released under the Elastic License 2.0. The core architecture uses Supabase (Postgres + Auth + Storage) as the backend, eliminating the need to run a separate database server. Deployment is straightforward: clone the repo, fill in a .env file with your Supabase keys and LLM provider credentials, and run either a one-command setup script (bash scripts/setup.sh) or docker compose up —build. Within minutes, the platform is running at http://localhost:8080.

The platform offers six major components: (1) Agent Chat, where users build an agent, wire up tools, and chat with it in-browser, with full request/response traces visible. Enabling Visual BI renders a chart from connected tables next to the answer, and agents can generate fully-editable PowerPoint, Word, or Excel documents — the latter pulling all rows with live formulas. (2) Swarm Canvas, a visual workflow designer built on XYFlow that lets users design multi-agent workflows as a graph and execute them end-to-end from the canvas, API, or a schedule. Runs checkpoint as they progress, so they survive restarts. (3) Knowledge Base / RAG, where documents are uploaded, chunked, embedded (using pgvector), and used to ground agents in custom data. (4) Data Sources & Catalog, supporting PostgreSQL, MySQL, Oracle, Redshift, Snowflake, Databricks, BigQuery, Synapse, Trino, Athena, and S3-compatible buckets. The Data Catalog crawls connected warehouses, infers schemas, profiles columns, estimates row counts, flags PII, and tracks lineage and usage across dashboards, prep flows, and metrics. (5) Business Intelligence Workspace, with drag-and-drop dashboards, 19 chart types, AI-assisted dashboard generation, click-to-cross-filter, drill-down, scheduled refresh with email reports, data alerts, row-level security, and export to Git repos. (6) Developer Workspace, a Python notebook environment on sandboxed server kernels with real CPython, pip install, and working imports of LangChain, LangGraph, and LlamaIndex — plus the ability to publish a notebook as a callable API.

Security and control are enforced throughout. Provider keys are encrypted and per-user; MCP server connections are supported. Secrets are stored encrypted and write-only, referenced as {{secret:NAME}} across the platform. The Developer workspace sandboxes Python execution so no provider key ever exists inside the notebook environment; model calls and knowledge-base searches are brokered by the platform. An MCP Builder lets users write MCP servers in Python and deploy them to the same sandboxed kernel; they register themselves, scale to zero by default, and support hashed API keys with expiry, per-tool, and per-IP limits. Changes to tool schema block calls until re-approval — a safeguard against unauthorized changes.

Observability is comprehensive: an execution trace inspector shows every tool call, token, and cost; an audit trail logs who did what (model calls, dataset queries, dashboard views, catalog crawls) with configurable retention; and admin-only spend analytics break costs down by user and IAM group. Web search and browsing are built in, using free DuckDuckGo by default or Firecrawl (or Brave / SerpAPI / Tavily / ScrapingBee per agent) for full-page reads, with SSRF guards on all fetches.

The repo also supports a parallel hosted service, agentswarms.fyi, which shares the same UI but serves a different mission: a fully managed learning platform with guided curriculum, build-along labs, interactive notebooks, and certification — no setup required. The author retains the AgentSwarms name and hosted service, while the source-available repo enables teams and tinkerers to run their own instance. Installation guides cover macOS, Linux, and Windows, and production deployment docs describe paths for local desktop, single cloud VMs, autoscaled VMs behind a load balancer, Cloudflare Workers, and Kubernetes — all with TLS, scheduling, health checks, and backups.

AgentSwarms addresses a common friction point for enterprises and teams: the need to run agentic AI infrastructure without outsourcing data or model calls to a third party. By using Supabase as the backend (eliminating the need for separate database infrastructure) and bringing-your-own provider keys, the platform ensures that data stays in-house and organizations retain control over their LLM consumption and costs. The Elastic License 2.0 enables commercial use while keeping the source code available.

The platform bundles several layers — agent orchestration, multi-agent swarms, business intelligence, data integration, and a sandboxed developer environment — into a single deployable stack. This breadth may appeal to teams that currently assemble these components separately or rely on hosted SaaS vendors. The visual swarm canvas and checkpointing mechanism (runs survive restarts and deploy cycles) are geared toward production use, while the Developer workspace and notebook-to-API publishing path cater to data teams and ML engineers.

The hosted variant, agentswarms.fyi, signals that the author is pursuing a dual-mission strategy: the source-available repo targets organizations that want to self-host, while the hosted service focuses on learning and managed deployment. This separation is explicit in the documentation and avoids confusion about which product is which.

No comments yet. Be the first to share your thoughts!