BYOK AI: Bring Your Own API Keys, No Model Markup

Last week some of you found our bugs before we did. A reply stopped after 26 seconds of work with "that took more than 30 minutes, so I stopped". A message sent while the connection dropped never arrived. One of our own octos checked whether a file existed 101 times in a row.

Since our October 20 newsletter we've shipped 90 bug fixes across Octomind Cloud, the panel and the Octomind CLI, along with features we'd wanted for a while. Model calls now cost exactly what the provider charges, you can bring your own API keys (BYOK AI), octos can search the web, and messages you send while an octo is busy wait in a queue on our side instead of getting lost.

Now we'd like you to try to break it.

Tell us what still breaks

Tell us what broke, what confused you, and what you expected an octo to do that it didn't. Post in Discord, reply to our newsletter, or open an issue on GitHub for the CLI. A screenshot and the time it happened is usually enough for us to find it.

If you have an Octomind Cloud account, this week's newsletter has a thank-you for that in it. It's in your inbox and on the panel's News page.

Models at the price the provider charges

Until yesterday, a model call cost the provider's price plus 5%. That 5% is gone. Every model call and embedding now costs exactly what the provider charges for it.

What we charge instead is 3% when you add credits, shown before you pay. Add $100, your card is charged $103, and $100 lands in your wallet. Plan prices don't change, and model use inside a plan is at provider price too, so the same allowance now covers about 5% more of it.

Prices that were already fixed stay as they were: image and video generation, voice, web search, machine time and storage. Everything is on the pricing page.

BYOK AI: bring your own API keys

If you already pay a model provider (startup credits on Anthropic, a company OpenAI account, an OpenRouter balance), you can now use that key in Octomind Cloud. Open Settings → Account → Your provider keys and add it. Eleven providers work today: OpenAI, Anthropic, Google AI Studio, OpenRouter, Ollama Cloud, Together AI, DeepSeek, Z.ai, Moonshot AI, Alibaba Cloud and Cerebras.

• The provider bills the tokens. We charge $0.05 per 1,000 requests, so a hundred requests cost half a cent.
• A model your key covers runs on your key, marked "your key" in every model picker. That includes models we otherwise sell on credits only.
• Several keys are tried top to bottom for the models they cover, and you can reorder them.
• Your requests never fall back to our keys. If your key fails, the request tries your next key, never ours, so a dead key shows up as an error instead of on your Octomind bill.

Keys work for chat models on paid plans. The panel shows each key's last characters and when it was last used; the key itself is stored encrypted and never shown again.

Octos can search the web

On paid plans, octos now have a web search tool. Ask about something newer than the model's training ("what changed in the latest PHP release?") and the octo searches, reads the results and answers from them. Each search costs $0.005, and only searches that return results are charged. Machines pick the tool up the next time they start.

It pairs well with a routine. A prompt like this, on a weekday 8:00 schedule, gives you a morning brief that is actually from this morning:

Search the web for news from the last 24 hours about Acme, Globex and Initech.
Give me five lines, newest first, each with its link. If there's nothing new, say so in one line.

Keep writing while your octo works

Send a message while an octo is still working and it waits in the chat with a dashed edge and a queued tag. Keep adding; messages queued together go out as one when the octo is free. ✕ takes one back, and Stop and send now stops the current reply and sends your queue right away.

The queue lives on our side, not in your browser. Close the tab, switch to your phone, let the laptop sleep: your messages still reach the octo when it's free, and every device shows the same list. A conversation holds up to 20 queued messages. Text you typed but didn't send stays in the box when you come back.

A dropped connection no longer loses the reply

Your phone goes into a tunnel halfway through a long answer. The reply keeps running on the machine, and when the panel reconnects it picks up where it left off: what you missed arrives once, in order, and the rest streams live.

That was also the cause of the false "that took more than 30 minutes" stop. A page that reconnected could land on a different one of our servers. The old server waited on a connection that had gone quiet and gave up; the new one never took the turn over. Now the new connection adopts the running turn.

Our own deploys don't cut replies anymore either. Each release starts beside the one that's running, takes the new traffic, and running conversations move across to it.

Two more things you'll see instead of a silent wait:

• Provider retries show up in the chat. When a model provider is down or rate limiting, the octo says so, for example Model provider unavailable (HTTP 503) — retrying in 60s, attempt 2/3.
• Loops end on their own. If an octo repeats the same tool calls and gets the same results 10 rounds in a row, the turn ends with a note naming what repeated, instead of paying for round 101. Change the limit under Loop stop in Config, or read how it works in the 0.61.0 post.

Watch your octo's team work

Bigger jobs get split: an octo hands the research to one specialist and the code to another. When that happens, a Team tab appears beside the chat. Whoever is working now is on top, finished runs fold below, and each run shows what it cost. Watch follows a live run's log; Log opens a finished one.

You can also choose the model for each specialist and each routine:

• Config → Specialist models runs a specialist such as developer:general on its own model: a strong one for code, a cheaper one for summaries. Specialists you don't set run your main model.
• Config → Capability providers picks the provider when a capability ships more than one, such as a browser through Playwright or Puppeteer.
• Every routine has a Model field. A morning digest can run on a small, fast model while the weekly report gets a frontier one.

New specialists for video and marketing

The community tap is the open registry of specialists an octo can hand work to. It now holds 157 of them, and the new ones this week are for people who sell with video:

• Clipper (video:clips) turns a podcast, webinar or stream recording into shorts. It transcribes once, ranks the moments that stand on their own, and waits for your picks. Each pick becomes a 20–40 second vertical clip with captions, a hook line and a cover.
• Footage editor (video:editor) cuts your own talking-head, demo or UGC footage into captioned shorts: silences and flubs out, reframed to 9:16, loudness levelled. Hand it last month's winning ad and it makes hook, length and aspect variants.
• Thumbnail designer (video:thumbnail) builds thumbnails from real frames or a generated scene, and sets the text itself instead of asking an image model to spell it. You get three variants for YouTube's Test & Compare, plus Reels and TikTok covers.

The tools behind them run on the machine with no API key: word-timed captions, silence cuts and reframes, and thumbnail text measured to fit. The video publisher now uploads and schedules to YouTube as well as TikTok and Instagram, through an Upload-Post account.

Decks and documents come out as files you can send. The presentation designer and the pitch-deck specialist render slide decks to PDF and PowerPoint, the proposal writer exports Word and PDF, and the résumé tailor renders ATS-friendly résumé PDFs. The writing specialists now pick a format by search intent (how-to, comparison, alternatives page, data study and more) and ask you for first-hand experience instead of inventing it.

Ask an octo in plain words ("cut this webinar into shorts") and it can hand the job to the specialist; the Team tab shows it working. On your own machine, run one by name: octomind run video:clips.

Open files right from the chat

• File names in a reply are links. Click report.pdf or cover.png and it opens in the viewer.
• Images and videos an octo generates are kept in Artifacts by default, so finished work stays with your account. Drafts it makes along the way stay on the machine.
• Opening or playing a file wakes a sleeping machine, so you don't have to start it first.
• Videos up to 256 MB download from the machine. The limit was 32 MB.

Update a machine when it suits you

When a newer Octomind is out, a running machine shows Update available with the version. The restart takes about a minute. It keeps your files, the shared folder, and every conversation, octo and routine. It resets programs installed outside your home folder (with apt or into /usr/local) and anything running at that moment. A stopped machine updates on its next start, with nothing to click.

The terminal drawer runs more than one shell now. New shell adds a tab, so a dev server can run in one while you work in another.

New models

Each has its own page on the models page, with how it compares to the models around it.

Smaller things

• A mark on octos with a new reply. The sidebar marks any octo that answered since you last opened it.
• Your invite link has a QR code on the referrals page.
• On phones, the chat fills the screen, and Return adds a new line on a touch keyboard; send with the button.
• Long step lists close from the bottom. An open list of tool steps ends with a Collapse button, so you don't scroll back up to close it.
• Config opens without a running machine, specialists and tap options included.
• Machines run Octomind 0.62.0 from their next start. The 0.62.0 and 0.61.0 posts cover what changed inside.

One we know about

When a model call fails partway through a reply, the conversation can stay busy for about a minute before it takes new messages, and Stop doesn't free it sooner. We're fixing that next. If you hit anything else, tell us. That's what this week is for.

— DK