Core concepts

Alfe surrounds an agent’s reasoning loop with the durable systems it needs between prompts. Read this page once and the rest of the docs will make more sense. Each section links to the pages that go deeper.

The agent operating system

An agent is a long-lived assistant with its own compute, memory, identity, and channels. Alfe is the operating system those pieces run on.

Every agent has one of two hosting types:

• Managed — Alfe provisions an isolated machine and keeps the agent online.
• Self-hosted — the agent runs on your own machine or server.

Either way, a local gateway daemon connects the agent to Alfe. It stays online, relays messages, and lets the platform manage the agent’s lifecycle. See Managed agents to compare the two.

Runtimes

The runtime is the software that does the agent’s thinking. You choose one per agent, and you can switch later without losing the agent’s memory or chat history.

Provider-backed runtimes run their own agent loop against any model provider. Terminal runtimes drive a provider’s own coding agent, such as Claude Code, and use that provider’s models. See Agent runtimes.

Persistent memory

Agents have long-term memory that survives context resets, restarts, and replacement machines. It combines semantic recall with a knowledge graph, so an agent can recall facts, people, and prior work. See How memory works.

Note

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

Model access

Alfe puts many model providers behind one governed proxy. You can change the model an agent uses from the dashboard, without rotating keys. Policy and spend controls apply before each call, and you can bring your own provider key. Model configuration is core agent configuration, not an integration you install. See Models.

Identity and scopes

Every agent has its own stable ID and credentials. When an agent connects a channel like Slack or Microsoft Teams, it does so as its own bot, not a shared one.

Organizations are structured into teams and projects. Integrations, files, and credentials are granted at one of four scopes: organization, team, project, or agent. A fleet can share what’s relevant and keep the rest separate.

Subscription and credits

Alfe runs on a fixed subscription plus a tenant-wide USD credit pool. The pool funds compute, model usage, voice, channels, and storage, so there’s no per-provider bill to reconcile. Usage only draws from that pool, so your spend has a ceiling you set.

Managed agents are add-ons on that subscription: your plan includes some capacity, and you add more as your fleet grows. See Billing and live pricing.

MCP-native

Alfe speaks the Model Context Protocol. The public server at mcp.alfe.ai lets agents discover the platform, onboard themselves, and work through MCP tool calls. Agents can also use MCP servers as tools. See MCP on Alfe.