Wolli

Wolli is an open framework for building AI agents that grow around a single purpose instead of resetting after every request. Rather than treating an agent as a stateless API call, Wolli runs it as an always-on daemon on your own machine, one that remembers past conversations, reacts to outside events, and extends its own abilities over time. The pitch on the site is to create agents that grow around a purpose, and the design follows that literally, since the agent works out what it's for and then builds itself out to serve it.

The core idea is that an agent is a home directory of ordinary files that both you and the agent can edit. Everything lives under a folder on disk, and because the files are the interface, there's no separate registry to maintain. When you or the agent drops in a new capability file, Wolli discovers it automatically. That plain-files approach is what lets an agent author its own skills, tools, workflows, and integrations, since writing a new capability is just writing a file into its own home directory rather than editing the framework itself. Deleting a capability is just as direct, since removing the file removes the ability.

Memory is handled through a small set of frozen files that sit in the agent's system prompt. SOUL.md holds its identity, and the first line is its purpose, which Wolli writes together with you in the very first conversation. MEMORY.md keeps durable notes the agent maintains for itself, and USER.md records facts about you, the human it works for. Alongside those, each conversation is stored as an append-only JSONL tree that acts as the agent's lifetime memory. Nothing is rewritten, so context is reconstructed deterministically from the log every time, which keeps the agent's history auditable instead of hidden inside a black box.

Being always-on is central to how it behaves. Agents run as operating-system daemons, launchd on macOS and systemd on Linux, so they can respond to schedules and events while your machine is on. A daemon exposes an HTTP interface, and clients like the CLI and the TUI attach to it. Workflows bind to typed events, so when a Telegram message arrives it emits a message event that routes to the right session, and a separate workflow fires when the agent finishes to deliver the reply. That event model is what lets an agent wake on its own instead of waiting for you to open a chat window.

Safety and flexibility are built in rather than bolted on. By default the agent runs inside a sandbox, using srt, which wraps Apple Seatbelt or bubblewrap, with Docker available as an option for stronger isolation. That matters when an agent is writing and running its own tools, since the sandbox limits what it can touch. On the model side it's provider-agnostic. You log in over OAuth and can point it at Anthropic, OpenAI, or others, so you're not locked to one vendor and can swap models without rewriting the agent.

The people this is aimed at are developers comfortable on the command line who want a personal, persistent agent they actually control. It ships with built-in integrations like Telegram and a scheduler, and the rest loads from the agent's own home directory as it grows into the job. Because it's local-first, you host it yourself and supply your own model credentials, which is the cost of keeping the whole thing on your own machine rather than in someone else's cloud. In exchange you get an agent whose memory, identity, and tools are all files you can read, edit, and back up.

What makes the approach distinct is that the agent is expected to change over time. Most agent frameworks hand you a fixed set of tools and a prompt, and the agent stays whatever you shipped. Wolli treats the first conversation as the moment the agent decides who it is, then lets it write new workflows and skills as the work reveals what it needs. Because every change is a file on disk, you can watch that evolution happen and step in whenever you want.

Wolli is fully open source under the Apache-2.0 license, with the code on GitHub under opsyhq and published on npm by Opsy, Inc. It's a TypeScript and Node project that needs Node 22.19 or newer, and installation is a single command, npm install -g wolli, after which running wolli starts the setup and the first conversation. There's no paid tier to the framework itself, so the practical cost is whatever you spend on the model provider you connect. For anyone who wants to inspect or extend it, the whole thing is a Node monorepo, so the framework's own internals are as readable as the agents it runs.