Core concepts

This page introduces the core ideas behind Alfe: the agent OS, persistent memory, model routing, identity and scopes, and the credit pool. Read it once and the rest of these docs will make more sense. Each concept links out to the reference and how-to pages that go deeper.

The agent OS

Alfe is an operating system for agents. An agent is a long-lived assistant with its own runtime, memory, identity, and set of connected channels. The runtime can live on your own machine or server, or on a managed host that Alfe provisions and keeps running 24/7.

A local gateway daemon connects the agent to the Alfe cloud. It stays online, relays messages to and from channels, and lets the platform manage the agent’s lifecycle.

Persistent memory

Agents have long-term memory that survives across conversations. Memory combines searchable history with a knowledge graph, so an agent can recall facts, people, and prior work — not just the last few messages. Memory is scoped to the agent and can be searched and written through the Agent API.

Note

Memory is private to each agent. Because the Agent API resolves identity from the agent’s token, a call only ever reads or writes that agent’s own memory.

Model routing

Alfe exposes a single model endpoint that routes to many LLM providers at pooled rates. You can switch which model an agent uses without changing any integration, or bring your own provider key (BYOK) when you want to use your own account. LLM configuration is core agent configuration — it isn’t a separate integration you install.

Identity and scopes

Every agent has its own stable identity and its own credentials. When an agent connects a channel like Slack or Discord, it does so as its own bot — not a shared, global one.

Organizations are structured into teams and projects. Memory, files, and context can be scoped at four levels — organization, team, project, and agent — so a fleet of agents can share what’s relevant while keeping the rest separate.

The credit pool

Billing is built around a single, tenant-wide USD credit pool. That one pool funds everything an agent does: compute, model usage, voice, channels, and storage. There’s no per-provider or per-token bill to reconcile — usage draws down the shared balance.

Managed agents are add-ons on a single subscription: your plan includes some capacity, and you add more as your fleet grows. See the pricing page for current plans and rates.

MCP-native

Alfe speaks the Model Context Protocol. A public MCP server at mcp.alfe.ai lets agents discover the platform, onboard themselves, provision compute, install integrations, and manage tokens — all through MCP tool calls. See MCP on Alfe.