Get the weekly radar
One short email with the freshest products, every week.
Wolli agents write their own purpose, integrations, and skills instead of staying fixed forever.
Grounded in available product and source data
Fixed tools do one job forever, never more — Wolli is built around CLI agents that write their own integrations instead, starting empty and growing more capable at a defined purpose over time rather than staying static.
A new agent works out its purpose with the user in its first conversation, writing that purpose as the first line of its own SOUL.md file. From there, the agent is self-extending — the site's own example, an agent called 'scout' that needs GitHub issues delivered the moment they open, writes its own GitHub integration, the workflow that wakes it on that event, and the tool it uses to report back, rather than a developer building each of those pieces by hand. When an issue opens, the relevant workflow wakes the agent in a fresh session to triage, flag what matters, and report — and any routine the agent repeats often enough gets written down as a skill, authored by the agent itself, so experience compounds into reusable capability instead of starting from scratch each time.
Memory is handled deliberately rather than as an afterthought: sessions form an append-only JSONL tree acting as the agent's lifetime memory, with nothing ever rewritten and the latest leaf resuming by default. Curated memory files — SOUL.md, MEMORY.md, USER.md — get frozen into the system prompt each session and are maintained by the agent itself as it learns.
Safety runs through sandboxing by default, using Apple Seatbelt, bubblewrap, or Docker depending on platform, with access to the real machine sitting behind an approval-gated escalation rather than being granted automatically. Model choice isn't locked to one provider either — multi-provider support via OAuth login covers Anthropic, OpenAI, and others. What the page reviewed doesn't show is any pricing information at all, consistent with Wolli being distributed as an installable, open CLI tool via npm rather than a hosted paid service.
Not necessarily — the site's own example shows an agent writing its own GitHub integration and the workflow that wakes it, once it determines that's what its purpose requires.
No — by default an agent is boxed in, using Apple Seatbelt, bubblewrap, or Docker depending on the OS, and touching anything outside that box is a deliberate step that has to be explicitly approved rather than something that happens by just running a command.
No — every session gets appended as a new leaf onto the same memory tree rather than overwriting anything that came before, so 'picking up where it left off' by default just means resuming the latest leaf, not starting over from nothing.
No — OAuth login is what makes the provider swappable: signing in covers Anthropic, OpenAI, and others without hardcoding an API key to one vendor ahead of time.
The page reviewed doesn't show any pricing information, consistent with it being distributed as an installable CLI tool via npm rather than a hosted paid service — worth confirming directly for any usage costs tied to the AI provider chosen.
