askalf

a self-hosted AI workforce platform — early access

A fleet of AI agents that triage, build, audit, and ship — on your server, under your control, with a human in the loop where it counts.

open-source building blocks — dario · hands · deepdive

01 · what it is

A workforce, not a chatbot

askalf runs a fleet of AI agents as a 24/7 workforce. An orchestrator takes work in — tickets, email, chat, schedules — triages it, and dispatches specialists: developers that open pull requests, auditors that verify findings before reporting them, monitors that watch production, a finance agent that sends the invoices. Anything risky stops at an approval queue and waits for you. It all runs on your own server — your data, your credentials, and your repos never leave your infrastructure.

02 · how it works

Intake to shipped, with you in control

01

Work comes in

Tickets, email, Discord messages, schedules, or events from your own systems land in one queue.

02

The orchestrator triages

Alf routes each item to the right specialist with context, tools, and a budget — and files what it can't handle for review.

03

Specialists execute

Agents work with real tools — shell, git, headless browsers, computer use, research, email — not just chat completions.

04

You approve what matters

Destructive or outward-facing actions wait in an approval queue. Budget gates cap daily spend. Everything is logged.

03 · features

Built for running real operations

Self-hosted

Deploys with Docker Compose on a server you own. Your data, credentials, and code stay on your infrastructure — no phone-home, no telemetry.

Runs on your Claude subscription

Model traffic routes through dario, the open-source router, so the fleet can run on a Claude subscription you already pay for — no per-token markup.

Human approval queue

Agents propose; you approve. Anything destructive or outward-facing — sends, deletes, deploys — waits for a human decision.

Budget gates

Hard daily spend caps enforced on every dispatch path. An agent that hits the gate stops — it doesn't quietly keep burning.

Discord control plane

Watch progress, approve actions with a button, and message your fleet from the same channels you already live in.

Persistent memory

Agents share long-term memory — decisions, gotchas, and context persist across sessions, so the fleet learns your stack instead of re-asking.

Real tools

Headless browser sessions, cross-platform computer use, cited deep research, email in and out, git and pull requests — the tools real work needs.

Observable end to end

Every execution, tool call, and dollar is recorded. A public status page covers the platform's own services.

04 · proof

It already runs a real studio

askalf isn't a concept demo. The platform runs Sprayberry Labs, a working software studio, in production today — the fleet triages the inbox, builds and reviews pull requests, audits client code, watches production, and sends the invoices. Early access ships the same platform we run ourselves.

05 · open source

Built in the open

The building blocks are MIT-licensed and on GitHub today — each one useful on its own, battle-tested inside the platform.

dario router

Local LLM router with OAuth multi-account pooling — run agent traffic through the Claude subscription you already have.

hands computer use

Cross-platform computer-use agent with the full action set — click, type, zoom, scroll — plus an audit log of everything it does.

deepdive research

Research agent that plans, searches, fetches, and synthesizes into cited markdown reports.

agent fleet node

The WebSocket bridge that turns any machine — Windows, Linux, macOS — into a fleet node with one install command.

browser-bridge browser

Hardened headless Chromium in Docker with session isolation and a health endpoint — the fleet's shared browser.

claude-sync portability

Pack a Claude Code session into a single portable file and resume it on another machine, byte-identical.

more at github.com/askalf — pgflex, redisflex, git-providers, install-kit, and friends.

06 · faq

Questions, answered straight

When can I get it?

Early access opens in waves from the waitlist — small on purpose, so each deployment gets real attention. Joining the list is the only queue; no dates we can't keep.

Is it open source?

The building blocks are — dario, hands, deepdive, agent, browser-bridge, claude-sync and more are MIT-licensed on GitHub. The platform itself is not open source today; it ships self-hosted, so you can see exactly what it runs, where your data goes, and what leaves your network (nothing).

What do I need to run it?

A Linux server (a small VPS is enough to start), Docker, and a Claude subscription or API key. The fleet can also enroll your existing machines as worker nodes via the open-source agent.

What does it cost to run?

Model usage can ride a Claude subscription you already pay for (via dario), and infrastructure is whatever your server costs. Platform pricing is announced with early access — we'd rather show it working than pre-sell it.

Is it safe to let agents act?

The platform is built around not trusting agents by default: an approval queue for destructive and outward-facing actions, hard daily budget gates, allowlisted email recipients, and a full audit trail of every execution and tool call.

Does it really run unattended?

It runs a real studio 24/7 today — with a human approving the actions that matter. Full autonomy is a dial, not a switch: you choose what the fleet may do alone and what waits for you.

Run your own workforce

Join the waitlist and get early access when your wave opens.

get early access