Concepts
The ideas that make eidan a finished agent rather than a framework. eidan is built as plugins
on the matbot runtime — a thin, isomorphic-TypeScript
agent engine — and adds the product layer: relational memory, interop, jobs, auth, and deploy.
Postgres (the eidan schema) is the source of truth.
Relational memory
Memory isn't a text file or a vector blob — it's a relational Postgres knowledge graph. Conversations, messages, events, knowledge, notes, decisions, jobs and artifacts are rows you can query, link, export, and back up. Persistence is keen: an inbound message hits the store before the model is called, and messages are append-only. Knowledge is Obsidian-style linked, not flat, so loading it is about relationships, not string search alone.
Plugins & the service model
A plugin is one TypeScript module exporting const plugin. It extends the host through matbot's
seams — tools, hooks (screen / contribute / toolcall / toolresult / followup), a
StorageBackend / KnowledgeIndex, providers, frontends, and a service registry.
Cross-plugin collaboration goes through the registry keyed by interface name (e.g. EidanMemory,
JobHandlers, Notify), so bundles compose without hard dependencies. A plugin can ship a whole
vertical feature — backend, data schema, behaviours, an admin screen — in one folder.
Agents & triggers
Beyond the default assistant, you define your own agents: each a persona, its own model, and one or more triggers — a schedule, a sensor, a webhook. Agents run under your identity with your tools, can relate to each other (delegate down, escalate up), and record every run so their activity is observable.
Jobs & delegation
Work you hand off lands on the eidan.jobs work-queue, routed by capability kind rather than
by machine. A node that serves that kind claims it, runs it, and reports back. This is how a chat
request to "fix this bug" becomes a background coding job the Sage bundle picks up and turns into a
pull request.
The vault
Third-party credentials live in a per-user encrypted vault (envelope-encrypted at rest, master key in the environment only). The agent uses them by reference — a key is never pasted into a prompt or handed to a model. A separate LLM-free write path lets a settings UI store secrets without the model ever seeing the value.
Interop — MCP, A2A, AG-UI
eidan speaks three open protocols on three boundaries: MCP (expose its tools to external agents, and call external MCP tools), A2A (other agents can delegate to it), and AG-UI (the wire between the agent and the web chat). Interop is in and out, so eidan plugs into an existing stack rather than replacing it.
Runs anywhere, multi-node
eidan runs on one node or many — a laptop, a Raspberry Pi, a Fly app — all sharing one Postgres. Concurrency and single-owner work (schedules, queues) are coordinated through the shared database, so you can scale out horizontally without a second source of truth.
Ready to try it? Head to Getting started, or browse Use cases.