One platform.
Two tiers.

Your agent chat lives permanently on the left rail. Pick any agent you've sprouted and start a conversation, or open a Conference to coordinate multiple agents on the same workflow at once. The right side shows your agent roster, cost tracking across whatever providers you're using, models currently loaded, and a live overview of your system's state.

It's the screen you'll spend the most time on, and it's built so you don't have to leave it for routine work.

Built for developers who want hands-on oversight of agent-initiated coding work. The Code bench is where your agent edits, runs, and tests code with you watching every move. It works today.

Honest framing: we aren't claiming superiority over dedicated coding agents on this surface yet. It's a focus area on our roadmap, and like everything else on the platform, federation means we learn from real participant use what to prioritize first. If something's missing or rough, the fastest way to fix it is to use Code, hit the rough edge, and Report it. The signal flows back into the build queue.

Schedule prompts to fire at your agents on a cadence. Daily summaries at 8am. Weekly competitive scans. A 4-hourly health check on a long-running build. Anything you'd otherwise have to remember to ask for.

Crickets are quiet, persistent, and accumulate value the longer they run. Every fire and result is captured in your agent's memory, so the system gets sharper at the recurring work over time. Routine prompts route to a cheaper model class by default — your premium model for the hard stuff, cheap models for the chirping.

Dynamic MCP surfaces every third-party platform you've connected and gives your agent supervised access to those accounts — Slack, Linear, GitHub, Notion, your CRM, anything that speaks the Model Context Protocol. Your agent walks you through each integration; you approve scopes; the connection lights up.

Honest framing on data: the third-party services on the other side of these integrations see exactly what you send to them — same as if you were using their app directly. We don't control that, and we won't pretend we do. What we do control is what your agent learns about your work in those services, and that part stays sovereign on your devices. The net result is more control over your data than the default way of using any of these tools, not less.

Marketplace is where your agent surfaces tools, integrations, skills, and bundles based on what you've actually asked it to help with. Ask your agent what's available; it goes and finds candidates ranked by federation telemetry from real participant use, then surfaces the top matches for your approval. No purchase, no install, no permission grant happens without your explicit consent.

One deliberate constraint: unlike siloed systems, you can't just grab arbitrary code off GitHub and bolt it onto your agent. This isn't a design flaw — it's intentional. Everything that runs through the marketplace is signed, audited, federation-ranked, and sandbox-tested. Anything else is a security incident waiting to happen.

If there's a tool or skill that already exists somewhere else and you want it inside Alpenglow, the right move is to reach out to the developer and tell them to publish it here. They'll thank you — for the first time, they can monetize the work directly. (See the marketplace economics for how that works.)

Hard rule for publishers: do not submit copies of code you didn't write. The audit pipeline checks every submission for authorship. Attempting to monetize someone else's work will get you locked out of marketplace submission. Permanently.

The Vault holds everything the platform itself needs to run — your agents, your memory, your conversations, your skills, your cricket history — encoded with substrate- native compression and encryption that's proprietary to us and not available anywhere else.

It's also yours to use directly. Drag and drop any file anywhere on your machine into the Vault and it gets compressed, encrypted, and stowed under the same architecture. Free up space on your devices by storing things in the most secure on-device file system that exists. Courtesy of Alpenglow — no charge for the storage, no upload to our servers, no third-party cloud.

Pair every device you run Alpenglow on — Mac, iPhone, iPad — into one trusted set. Surfaces talk to each other directly, device-to-device, end-to-end encrypted, with no relay through our cloud, no central key escrow, no man in the middle. State syncs in real time across whatever's online; the rest catches up when it comes back.

Lose a device, revoke it from any other one in the set; the remaining devices keep running. Add a new one, walk through the pairing flow, and your agents and memory are there in minutes. Your trusted set is yours; we don't hold the keys.

Your agent's memory surface — the working state and accumulated history that makes it sharper at your specific work over time. Browse what it has learned, search across years of conversations, see how procedures and preferences are organized, prune anything you'd rather it forget.

The deeper architectural argument lives on its own page (how memory works); this is where you actually look at what's there. No vector database, no embedding similarity search, no probabilistic recall. Exact memories, exactly retrieved.

Community Voice is the active wishlist — the things participants are openly asking us, or asking developers, to build. Skills that don't exist yet. Integrations that should. Dashboard modules someone will pay for once they ship. Workflows worth automating. Anyone can contribute, anyone can upvote, the federation surfaces the requests with the most weight.

Developers, this is your free leads list. Every entry is a participant openly saying "I'd pay for this if it existed." We can't make everything everyone wants — but the active list is the community asking you to monetize the gap. Build it. Submit it. The marketplace audit pipeline takes it from there. Federation ranks it. Participants install it. You get paid.

Provider keys, cost defaults, model preferences, recovery phrase, account wipe, federation participation toggle, data-export — all the controls one tab away. Defaults are sane, so you usually don't need to touch much. When you do, everything's where you'd expect it.