What we shipped, and why.

Getting started used to walk some people through as many as eleven screens, and ask for a card before showing anything. It is now six screens at most, with one decision on each, and the steps that were cut are still available later from Apps, Settings and Home.

The heart of it is a short map of how your work runs: for each kind of work you do for clients, what comes in, what goes out, how often, and who signs off. Mycel fills in a first guess from your website, anything you uploaded, and the description you gave. You correct it with a few taps, not a form. It starts from your firm as it is, rather than from a generic picture of your trade.

While you answer, Mycel drafts one real piece of your work, from a file you dropped in or, if you did not, from a clearly labelled sample client, and you watch it being written. You see that draft before the plans. You can still start your seven-day trial right there, card up front, for the $299 a month Starter plan, or choose "Not yet — show me the work first" and look around. Starting the trial stays one click away on Home.

Sign-up is email and password or Google. We ask for your website on the first screen of setup, and use it to set your brand colours and to know what you sell.

When a run failed, you used to get a raw error message. Now a failed run says, in plain words, what happened (it stopped responding, it never started, a rule blocked it, a service we rely on was down) and what to do next. The technical detail is still there, folded away, if you want it.

If you leave a page open while we ship an update, it now reloads itself onto the new version the next time you use it. Before, it showed "Couldn't load this".

You can also disconnect any app you have connected, like a mailbox, a bank feed or your accounting system. Before, a connection could be replaced but never removed.

A finished piece of work used to open in a viewer. Now every link to a deliverable opens the studio, and the studio is where you change it.

Click into a document and type: headings, paragraphs, list items, table cells. A PDF, including one you uploaded yourself, is edited on its own pages rather than turned into something else first. Spreadsheets recalculate as you change an input, and you can insert or delete rows and columns, sort, fill down, and paste a block straight from Excel or Sheets. CSV files, which used to be read-only, edit like a one-sheet workbook. HTML and plain text get a proper editor with find and replace, and images can be rotated, flipped and cropped. Your edits collect into one change set, and saving makes a new version.

If you would rather describe the change, ask. You can pin notes on the parts you mean and send them all as one request, then watch the rework happen. The chat keeps every conversation you have had about a document, so you can come back to an earlier thread. Versions are named for what they are, side-by-side compare and restore are one click, and the whole version downloads as a zip. From the Download menu, a written document can be exported as a branded PDF or a Word file, and a CSV as Excel.

Beside the actions, the studio now estimates how likely this client is to accept the draft as it is, and why. It bases that on what has actually happened: how often your drafts go out unchanged, how this client has responded before, and whether the draft still has gaps marked to confirm. When there is no history yet, it says so instead of guessing.

Correcting a draft used to fix that draft and nothing else. Now, when you edit a deliverable in the studio or ask for a rework, the change becomes a rule for that client's next runs. It applies to that one client, not your whole book, so a preference one client has does not leak into work for another. After you save, the studio tells you what it learned, and for whom.

There is also a page for what Mycel has learned from you: every rule, where it came from, how often it has been used, and buttons to edit it or forget it. Each client shows how often you have had to correct its work, and every version shows anything unusual in it and the checks it went through, at the top of the studio and in the deliverables list.

The deliverables list itself is now a table, with search, filters for client, service and date, sorting, and saved views. Use the arrow keys (or j and k) and Enter to open one in the studio.

Find clients now opens on buying situations: who is likely to need a business like yours right now, why now, and the search that would find them. They come from what you sell and who you already work with, whatever your trade.

Any of them can become a watcher. Once a week it looks at public pages, starting with dated news, and every business it finds is added as a lead, with the exact sentence that pointed to it quoted so you can check it yourself. From a lead, "Find the owner" opens a search already filled in.

And if you already have thirty clients in a spreadsheet or a CRM export, you do not have to add them one form at a time. Paste the list. It copes with rows copied out of Sheets or Numbers, or typed one per line.

An approval used to show you a draft and a button. Now it tells you who it goes to, by which channel and at which address, and when it expires. Before anything is sent from the list, a confirmation names the recipient. If sending is not possible yet, the reason is shown right there instead of a greyed-out button. Editing the wording is a clear button, and saving an edit tells you it will shape that client's next draft.

Fifteen near-identical emails used to be fifteen cards. Now they are one decision. On any approval, "Stop asking me about these" lets that kind of work go out without you. It tells you exactly what it covers, with a daily limit (for example, up to 50 reminder emails a day), and how to undo it. The separate standing-approvals page is gone; what runs without you is shown on Home as a set of switches with those daily limits beside them. Anything high-risk always waits for you.

Your clients sign into the portal from a link, with no password. They see the work you have sent, and accepting it is one button. You can open the portal from a client's page to see exactly what they see.

The portal is now light and plain rather than a dark dashboard. It no longer opens on a row of numbers, empty sections are quiet, and it fits on a phone without squeezing your firm's name. It also shows what the client has received, not only what they owe.

When you ask a client for a document, they can drop the file in and send it. A note is optional. When they pay an invoice by bank transfer, they can press "I've sent the transfer" and add a reference, and the reminders stop. When you record a payment, their view shows it settled straight away.

You can also serve your site and portal on your own domain. Claim it, add the DNS records Mycel shows you, and it handles the certificate and goes live.

There is now a Yearly option on the pricing page and in billing. It charges once for the year, at ten months' price: Starter is $299 a month, or $2,990 a year. The seven-day trial works the same way either way, with the card taken up front, and the pricing page says exactly what will be charged, and when, for the option you chose.

Starter is also where every business now begins. The separate $49 plan for finding clients is gone, and finding clients is part of Starter alongside the delivery work.

Signing up now builds your business first and asks for a card second. You describe what you sell, Mycel sets up the work, the schedule, the accounts and your site, and you land inside the real product with everything in place. Nothing is a demo.

What you cannot do yet is run it. Until you start your seven-day trial, no jobs go out — and a calm banner at the top of every page says so, in one line, with one button. No dead ends, no guessing why nothing happened, no surprise charge. You look around as much as you want, and the moment you are ready, one click starts the trial with the card taken up front, so day seven is an ordinary renewal rather than a fresh decision.

If a plan ever lapses, the same banner appears and says the same true thing: everything you built is exactly where you left it, and restarting picks the work back up, schedules and all. We never delete your businesses, your knowledge, your connections or your history over a payment.

Finding new clients used to work, then not, then work again, with no way to tell which you would get. The reason was the source: it let a couple of searches through per address, then started handing back stripped, empty results while still looking like a success. On a bad afternoon that was fourteen runs in a row that found nobody.

Finding now leads with a source that does not rate-limit us that way — the same one that already enriches the people it finds — and only falls back to the throttled path when it has to. In testing it returns real people with real profiles on the first try, every try. The run before this change: didn't finish. The run after: done, with people on the wall.

This is the whole point of the Find plan, so it had to be boring and reliable rather than lucky. It is now boring.

Find clients used to open on a form. Two of them, actually — start a campaign, set up autopilot — stacked tall down the page, with the real thing, the people we found for you, pushed to the bottom. That is a setup screen wearing the clothes of a workspace.

Now the people come first. The page opens on who is in your pipeline, the way a customer book should. Starting a campaign and setting up autopilot are still right there, one button each, folded away until you want them. Nothing was removed — the composer, the weekly autopilot, the whole sequence builder are all where they were, just no longer in front of the people they are about.

Inside a campaign, the pipeline already reads like the CRMs you know: a board you drag people across as they move from found to replied to booked, a table when you want columns, faces when you want to remember who they are, with filters and saved views that follow you between devices.

A real invoice chase wrote its answer five minutes in. Everything the task asked for, correct. Then it spent twenty more minutes reading the same policy file again until a time limit killed it.

The run was recorded as failed. The customer would have been told their invoice chase did not work, while the finished answer sat in a sandbox we deleted a moment later.

The problem: the run ended when the agent said it was done. That is not the same question as whether the work is done. So the harness watches the file instead. When the result file is there and matches what the task promised to produce, the run ends. Nothing left to pay for.

Free runs also got faster and cheaper. Free was capped to the cheapest model. That model kept getting one tool call wrong and repeated the same broken call eleven times without ever getting it right. It is 4x cheaper per token and far more expensive per job, because it loops until a ceiling stops it — and the ceiling is where the bill lands. Free now reaches the standard model. Spend is still capped in dollars, which was the control doing this job all along.

Found by an audit, before a customer found it.

Every API route scoped uploads and answers to the business that wrote them. The one place that puts those files in front of the agent did not. So two businesses running the same wedge shared a pile: uploaded documents, intake answers, prices, policies, and past corrections.

It was not only a disclosure. Intake answers use the same filename for everyone, so the last business to answer a question quietly overwrote everyone else's answer. Your agent could quote a stranger's late fee to your client, and nothing anywhere would say why.

Reading knowledge now requires a business id and fails closed. An unlabelled row belongs to nobody rather than to everybody.

Also: what we charge you was measured wrong, by 15-60x. The meter used hardcoded prices from a different model provider than the one we run. The free tier's "$2 a month" of usage was really about three cents. It reads the real price table now — which is the same number the plan ceiling enforces, so the limits mean what they say.

We had 700 lines of machinery for turning your edits into rules the agent follows. Nothing called it. What actually ran was: mount every file for this wedge and hope.

Now, when you edit what the agent proposed before approving it, your version is captured and turned into a rule. The next run gets the rules that matter for that client and that job — not everything ever written, which grows without limit and drowns the relevant thing.

Three decisions we would defend:

• A rule about one client beats a general rule, because the exception is the expertise.
• Auto-approvals do not count as learning. Nobody looked at them. Counting them would let a wedge that approves everything report itself getting better while getting worse.
• If the agent runs out of room for rules, it says which ones it dropped instead of silently cutting the list.

Your own words decide it, not the agent's paraphrase of them.

An internal, unlimited plan was applied to everybody by default, including accounts on our hosted product, and that plan has no seat limit, no job limit, and no spend ceiling. We bought the tokens. Nothing capped them. The pricing page was decoration.

Hosted signups now start on Free, so nobody acquires a bill by accident.

A connected app now means a connected app. We marked an app connected the moment you opened its login screen. Close the tab without finishing and it still said connected — with a green pill — so the agent was told a tool existed, reached for it, and failed at the moment it tried to do real work. Connected now means the provider confirmed it, and it clears when they stop confirming. Half-finished is its own state, so you get "finish connecting" instead of a lie.

The LinkedIn executor ignored which action it was asked to run. Endorse a skill, view a profile, anything at all — it sent a direct message. So the card you approved named one thing and a different thing happened. That makes the approval a decoration, which is the one thing it can never be.

There are seven named LinkedIn actions now, and anything outside that list is refused by name rather than guessed at. The sequencer and the executor read the same list, so they cannot drift apart again.

The outreach loop also runs end to end: find people, add them to a campaign, pace it, warm up, invite, message. People found this way are written into the CRM as records, and their employers as companies — the screens that have been rendering empty now fill.

Two rules worth naming. A campaign cannot grow after you approve it: the card said "N prospects", and a list that grows after your signature is a different decision wearing your consent. And a prospect added without copy parks rather than improvising words to send a stranger.

When the kernel starts, it looks for jobs that were interrupted and cleans them up. That is right when there is one process. We run several. So any container booting — a deploy, a scale-up, one health check restarting one task — marked every job in flight failed, across every customer, with "interrupted by a restart".

Found by running a real job twice and getting that error both times on a settled deployment. The second one was killed by our own probe container.

Cleanup now only touches jobs that have gone quiet for ten minutes. A live run writes constantly, so silence is the difference between a run whose process is gone and one another machine is still driving.

And there is a run page. A to-scale waterfall of what the agent actually did, drawn on the server so it works without JavaScript, then live over the stream: tool results, cost per step with the model and token count behind it, and where it drifted. Failure is treated as the normal case, because it is — "ran out of time" and "an approval was never answered" used to read identically, and now say which. The step where a run died stays amber and open, because nothing is running there.