Agentis · the open-source home for agents you own
Own the agents. Rent the models.
Agentis is local-first infrastructure for long-lived agents. Import the agents you already run, give them a permanent Brain of memory and skills, orchestrate them into self-healing apps that reach any channel — and swap the model underneath them anytime. Identity, memory, and reach belong to Agentis, not to whatever runtime executes a turn. 122 pages across 12 categories, researched against the live codebase.
Why it's different
Most agent tooling rents you intelligence: the model provider holds the memory, the skills, and the lock-in. Agentis inverts that. The durable half — who the agent is and everything it learns — is yours and portable; the model is a swappable tenant.
Runs on your machine with embedded SQLite and a bundled offline embedding model. Zero external services to start.
Import agents you already run in Claude Code, Codex, and Cursor — identity, memory, and skills become operator-owned data.
Durable memory, living skills, and grounded knowledge that compound across every run — separate from any model.
The Runtime Abstraction Layer moves an agent between runtimes without losing a memory, skill, or habit.
Typed workflow graphs that diagnose failure, repair the plan, and judge whether the goal was actually accomplished.
Agents author their own apps and datastores, then reach WhatsApp, Slack, Telegram, email, and the web.
The six primitives
The mental model is small on purpose: one durable substrate and six primitives. Every feature is one of these, or a composition of them.
Identity, sessions, runtime routing, and specialist execution.
SubjectA durable per-person actor that can wait for replies without spending tokens.
ConnectionChannels, integrations, MCP, A2A, and webhooks.
OrchestrationTyped workflow graphs, repair, replay, and delivery judgement.
ExperimentOutcome learning, trust controls, budgets, and governance.
InterfaceAgent-authored apps, collections, records, and AG-UI surfaces.
Every category
Read top to bottom — categories are ordered by how much they matter to a new reader, not by internal architecture layering.
Install, boot the runtime, and bootstrap your first agent.
02. Own Your AgentsBring the agents you already run. Swap the model, keep the mind.
03. Talk to Your AgentsChat, terminal, broadcast, and the live interaction feed.
04. The BrainDurable memory that compounds instead of resetting every session.
05. Reliable by DesignSelf-healing workflows that tell the truth about whether they worked.
06. Agents & CognitionHow a single agent thinks, and how many agents cooperate.
07. Apps & InterfacesDurable products agents build, staff, and operate themselves.
08. ReachOne agent, every channel — WhatsApp, Slack, email, MCP, and A2A.
09. Trust, Governance & SecurityBudgets, approvals, audit, and defenses that hold up in production.
10. The Agent-Native SDKOne typed surface agents operate as code.
11. ArchitectureThe durable spine under everything, for the technically curious.
12. ReferenceThe full tool registry, HTTP API, CLI, and glossary.