Getting started
Documentation
AgentSwarms is a platform for building agentic AI on your own data — agents and multi-agent swarms, the knowledge and tables they work from, the dashboards beside them, and the controls that make all of it safe to hand to other people.
New here?
Start here
Quickstart
Build your first agent, give it data and tools, run it, and read the trace.
Core concepts
Agents, tools, retrieval, swarms, memory and BYOK — what each really means.
Agent Builder
Every field on an agent, and what changing it actually does.
Install & deploy
Run the whole platform on your own infrastructure.
Find it by what you're doing
Get oriented
| Page | Covers |
|---|---|
| Dashboard | The screen you land on: quick actions, workspace stats, spend by person or team, activity, recent runs — and what to do first on an empty workspace. |
| Account | Your identity and credentials, what belongs to your account, and what happens to it when you leave. Spend caps live in Budgets. |
Build something
| Page | Covers |
|---|---|
| Agent Builder | Prompt, model, tools, knowledge, memory, guardrails, versions and export. |
| Agent Chat | Run agents, visual BI answers, PowerPoint/Word/Excel generation, sources. |
| Swarm Canvas | Multi-agent graphs: routers, loops, parallel work, approval gates, framework export. |
| Skills & Prompt Library | Reusable capabilities and prompts. |
| Developer workspace | Python notebooks with real kernels, model access and KB retrieval. |
Work with data
| Page | Covers |
|---|---|
| Data Catalog & SQL | Uploads, warehouse connectors, catalog, profiling, lineage, the SQL workbench. |
| Data preparation | Visual joins and transformations saved as repeatable, schedulable flows. |
| Knowledge Base | Ingestion, chunking, embeddings, graph search, and debugging bad answers. |
| Semantic Layer | Governed metrics so one question has one answer. |
| BI Workspace | Dashboards, filters, drill-through, alerts, sharing and embedding. |
Ship and integrate
| Page | Covers |
|---|---|
| Integrations | Connect providers, search, automation and data sources. |
| Models & providers | BYOK, the model registry, and choosing the right model per job. |
| MCP servers | Give agents tools owned by other systems. |
| Web embedding | Put an agent or dashboard on your own site, safely. |
| API & webhooks | Run swarms from your code: scopes, idempotency, signed callbacks. |
| Secrets | Store credentials once and reference them everywhere. |
Govern and operate
| Page | Covers |
|---|---|
| Access control | Users, groups, model rules, read-only sharing, invite-only signup and SSO. |
| Guardrails & PII | Input/output controls and personal-data detection or redaction. |
| Budgets & cost | Spend caps per user, group, embed or API key — plus retention windows. |
| Logs & traces | Read what an agent actually did, rather than what it says it did. |
| Analytics & audit | Usage, spend attribution and the audit log. |
Three things worth knowing up front
Prose and numbers are different problems. Documents go in the Knowledge Base to be quoted; rows go in the Data Catalog to be counted. Models are poor at arithmetic and excellent at sounding certain, so putting a number where only prose retrieval can reach it is the most common way to get a confident wrong answer.
The trace is the truth. An agent's explanation of its own reasoning is generated text and can be wrong. The trace is a record. Debug in that order.
Your keys, your data. Connect your own model providers and calls run under your account and your agreement with that vendor. See Models & providers.
When the docs do not have it
These pages describe the platform as it ships. For anything they do not cover — a question about your own setup, a bug, or a feature that should exist — there are three places, and which one you want depends on what kind of answer you need.
| Where | Best for |
|---|---|
| Discord | A question with no obvious answer, a sanity check on an approach, or anything where a conversation beats a written report. Fastest route to a human. |
| GitHub issues | Something reproducibly wrong, or a feature request worth keeping a record of. Include the version, what you expected, and what happened. |
| Contact | Anything you would rather not post in public — an instance-specific detail, a security report. |
Bring the trace