You've let the agents in. Nobody's watching the door.
Sound familiar?
- The briefLives in a chat thread nobody can find next week.
- An agent draftLooked great. Went live. Nobody checked it.
- A client's API keyPasted into a prompt on Tuesday. Still there.
- Eight toolsEight versions of what's true.
- A change to ad spendMade by a script. No receipt.
- What it costNo idea until the invoice lands.
And that's the visible part. The agent that guessed, because nothing was written down, is costing you more.
It's not that you don't care about the risk. It's that the tools ask you to choose: let the agents loose, or keep your hands on everything and lose the speed. You shouldn't have to pick. A person should decide what matters, and a machine should do the rest and show its working.
We hit every one of those running our own agency. So we built the thing we needed.
Agency Astro is a small marketing agency in Townsville. Over 2026 we replaced our project tool, CRM, client portal, wiki and approvals inbox with one shared record, then taught agents to work inside it. Hub is that system. Launch Astro is the name we're releasing it under, so the product has a home that isn't the agency.
What's actually behind it
And the honest line: as at the start of September, no job has run start to finish through the new run ledger. The board carries that as its highest-weighted open risk, in those words. It's the first thing we're building in public.
The hard part was never getting agents to do the work. It's seeing what they did, judging whether it's any good, and deciding what goes live.
Three steps from curious to running it.
- 01
Watch it get built.
Star the GitHub org and read the roadmap. Every row says what it's waiting on. When something moves, the board moves the same day.
Free, now - 02
Read the docs while the scrub finishes.
The architecture, the security model and the stack are written up here already. The public repo lands as a fresh-history mirror once every trace of client data is out of it.
Free, now - 03
Run your own Hub.
Clone it, point it at a Postgres and deploy it on Vercel. Add a repo of your own automations and skills. Your data, your keys, your machine.
First public release
One record. Four decisions. Everything else is a machine's job.
People make four decisions: what to do, whether the plan is right, whether the result is right, and whether it goes live. Machines do everything between those four moments. They can't hold keys, can't spend past their permission slip, and write down every attempt, including the refusals.
The loop
Signal
A source changes, or a person asks.
Interpret
Hub explains why it matters.
Task
The work gets an owner and an outcome.
Workflow
Jobs run in dependency order.
Review
Evidence and the exact version are checked.
Action
Approved execution produces a receipt.
Memory
The recap feeds the next run.
And back to 01
Step 07 is built and switched off. Nothing has produced a recap yet. We say so because the loop is the claim.
Eight areas. One database. No second source of truth.
Each card carries its live count off the board: capabilities built, partial and missing. The numbers move when the work does.
Projects & tasks
Boards, the task table and drawer, subtasks, comments, estimates and actuals, time tracking, My Tasks. The most finished area.
Clients & CRM
Directory, client record, contacts, leads pipeline, interactions ledger, tickets and enquiries, and the journey engine that writes follow-up tasks.
Client portal
What the client sees: track record, reviews and approvals, projects, reports, docs and account. Real records through row-level security, open since 1 September.
Knowledge & skills
The block-editor wiki, snippets, the client library (brand, voice, design system), and the skill contract that agents work from.
Agents & workflows
The approvals desk, the model resolver, the worker fleet, the run ledger and the gates. The area the whole platform turns on.
Signals & evidence
Site health, refusal-safe client numbers, the notifications hub, the ingest workers, and a freshness stamp on every displayed figure.
Security & access
Row-level security across every client-reachable table, the role model, step-up auth, secret custody, agent API scoping, the audit trail and the permission floor on approvals.
Platform & estate
The test battery, the parity harness, CI, replay tooling, the deploy paths, backups and alerting. Reported separately because no Hub user touches most of it.
Doing nothing isn't neutral. It compounds.
Every month the agents run unwatched is another month of drafts that went live unchecked, keys that sat in prompts, and spend nobody can trace to a decision. The agencies that get this right early will run more clients on fewer people. The ones that don't will find out from a client.
Imagine opening your desk to a review queue, not an inbox.
- 01Overnight work is staged, versioned and waiting for one yes.
- 02Every number on a report names its source and how old it is.
- 03A client's key never leaves the vault. A stolen slip expires in hours.
- 04What a run cost sits on the run, not on next month's invoice.
- 05The plain-language recap is filed under the client before you've read it.
That's not a someday. Most of it is code on main already. The roadmap says exactly how much.
Boring, portable parts. On purpose.
The record is the product. Everything around it is swappable, and the rebuild is choosing parts a stranger can run.
- Next.js · React · TypeScriptThe app, being rebuilt from the Vite + JavaScript original
- Supabase · PostgresThe record: tables, row-level security, realtime
- NangoConnection lifecycle for every platform Hub reads from or drafts into
- AirbyteScheduled pulls into a fenced landing zone
- n8nThe proven automations that pre-date the run platform
- VercelHosting, previews and the deploy path
Asking one agent gives you an answer. Wiring an organisation gives you a checked one.
Asking one agent, versus wiring an organisation of them. Named steps, run in parallel, checked by something that didn't write them.
- You ask → it works it out → you get an answerOne step, and whether it is any good is yours to spot.
- You ask → the work is plannedBroken into named steps before anything runs.
- Research, write and build, at the same timeSpecialists, each with one job.
- Someone separate checks itA checker that did not write the work.
- You approveNothing goes live until a person says so.
Every row public. Every row says what it's waiting on.
The board is the backlog authority for the repo: 436 rows across eight areas, each with a status, an impact and a named blocker. It's generated from the same file the agents work from, so it can't drift from the code.
Questions people ask first.
Yes. The platform core is genuinely open source, licensed AGPL so a hosted fork has to publish its changes. You run it on your own Postgres and your own hosting. The only running cost is whatever AI models you point it at, and the local-model floor is free. The skills, automations, workflows and presets built on top are a separate, proprietary layer.
Not yet. Agency Astro runs it in production, but the public repo is a scrubbed, fresh-history mirror that lands once every trace of client data is out. Until then, the roadmap and the docs are public and the board moves in the open.
Code that's merged with its tests running, held shut by a flag or a missing setting. It's real and you can't use it. We label it separately from 'written, not applied' (a migration no database has run) and from 'not built', because they're different distances from working.
Frameworks give you the node and the edge. They don't give you the record, the gate, the receipt or the credential broker, and those are where the risk lives. Hub is the operating layer around the agents, not another way to call a model.
Whichever passes the skill's tests cheapest. Frontier models for the hard reasoning, a local model as the free floor for triage and simple work. No skill is welded to one model, and each organisation brings its own key.
In your vault and one authorised runtime home. An agent never reads the vault. A task gets a signed, scoped withdrawal slip that expires within hours, and the broker does the transaction. Every attempt, refusals included, lands on the statement.
Consulting, and the content layer. The core is free and stays free. The skills, automations, workflows and dashboard presets we've built for marketing agencies are sold with the setup work, and the afternoon-install runbook is itself part of the product. Free users are the marketing. That's the trade, and we're fine with it.