StellarFirm is a new platform that gives founders and small teams a staff of AI assistants: a Coder, Marketing, an Accountant, Legal and more. You play the CEO. They split up the work, pass it to each other and stop for your Approve before anything risky goes out.
The promise is attractive, and the paperwork is unusually detailed for a product this young. It is also inconsistent. Its documents disagree on what needs approval, and our sums put credit prices well above model list prices. This review uses the vendor’s own site, docs, terms and privacy policy as they stood on 1 November 2026. We did not sign up or run a task, so treat it as a document review rather than a hands-on test.
Table of contents
- What StellarFirm Is, and What Is Actually Live
- The StellarFirm Team: Roles, Starters and Limits
- How StellarFirm Approvals Really Work
- Where the StellarFirm Documents Disagree
- StellarFirm Pricing: Plans, Tiers and the Launch Offer
- What a StellarFirm Credit Really Buys
- StellarFirm Data, Privacy and Security
- Who Is Behind StellarFirm
- StellarFirm vs Lindy, Sintra and Paperclip
- Should You Use StellarFirm?
- StellarFirm FAQs
- References
What StellarFirm Is, and What Is Actually Live
A Company of Assistants, Run From a Chat
The model is simple to describe. You write a goal in chat, a CEO assistant writes a brief and picks the right worker, and that worker acts on a private computer of its own. Work can also arrive through Slack, Telegram, Discord or Microsoft Teams.
The site calls these workers both AI Agents and Agentic Employees. If you want the wider idea first, our guide to AI employees and autonomous agents covers what this category promises in general.
A Very Young Product
The timeline is short. RDAP records show the stellarfirm.ai domain was registered on 2 October 2026. The terms and privacy policy are dated 14 October, the same day the premiere video went up on YouTube, where it had about 1,500 views when we checked. The blog’s first posts are dated 24 October.
We found no independent review. A Hacker News search returned no results, and the homepage’s roughly 100 directory badges sit beside a credit reading “Directory submissions done by submi.io”. Those are listings, not tests.
What You Can and Cannot Use Today
The vendor separates shipped features from planned ones more clearly than most young startups, which deserves credit. The roster below reflects the homepage and the compare page.
| Feature | Status the vendor gives |
|---|---|
| Web app with cloud workspaces | Available |
| Sixteen listed assistant roles | Available on the web |
| Desktop app for Mac and Windows | Early access, a pre-release |
| Android and iOS apps | Coming soon, not in the stores |
| Coworkers in one organisation | Coming soon, a later add-on |
| Marketplace of expert teams | Coming soon, nothing open yet |
| Orbit proactive checks | Coming soon |
| Desk pet hardware | Concept only, no unit built |
The StellarFirm Team: Roles, Starters and Limits
The Coder Is the Centre of Gravity
The vendor’s own instructions for agents call coding “the best fit today”, and most of its guidance backs that up. The Coder connects to GitHub, GitLab or Bitbucket through a token you create, works in its own workspace, runs your tests and opens a pull request.
The other roles are mostly drafting and preparation. Legal produces “drafts and checklists only, not legal advice”. The Accountant triages invoice mail and drafts reminders. Support prepares replies for you to check. That is a sensible boundary, and it also means a person with real expertise still has to read the output.
Starters for Unusual Trades
Each starter template bundles assistants, tools and prompts for one kind of business. They run from coding studios and marketing agencies to farms, university courses and mechanical design. The range is unusually wide, and it raises a fair question about how deep each role goes. A role that reads well on a landing page has still to prove itself on your real work, and we could not test that here.
Limits by Plan
Plan limits decide how large your virtual staff can be. Hobby allows 6 assistants and 2 teams, Plus allows 15 and 5, and Ultra allows 30 with unlimited teams. Custom roles need Ultra or Business. The bigger constraint for now is human: with coworker accounts still marked as coming, this is effectively one person managing many assistants rather than a shared workspace for a team.
How StellarFirm Approvals Really Work
The Promise Versus the Table
The homepage states that “every email, post and client proposal waits for your Approve”. The terms are more careful: nothing is sent or published until you approve it, but “request text may still go to a model provider before you approve it”. So approval controls what reaches a recipient, not what reaches an AI company.
The most useful single page is the approvals table, because it lists what does not ask. Assistants act alone on anything they consider making or changing, and some of those actions are visible outside your private workspace.
| Step | Does it ask you? |
|---|---|
| Pushing a branch or changing code on your repository | No |
| Opening a pull request or an issue | No |
| Saving a new file, for example in Google Drive | No |
| Using your credits on normal work | No, a budget stops it |
| Sending mail, messages or invites | Yes |
| Deploying or deleting | Yes, unless you enable auto-approve |
| Posting on social media | Blocked until you change the default rule |
| Spending your money | Always, with no switch |
The Parts That Hold Up Well
Several rules are stronger than the marketing needs them to be. Spending real money always asks, even from a routine. Auto-approve starts off, covers only production actions and deletions, and needs a confirmation to enable. Routines cannot skip an Approve, and running out of credits never approves anything. Rejecting a step cancels it.
The Line Worth Reading Twice
One note matters more than the rest: “Once you approve a kind of step in a chat, the same assistant does the same kind of step in that chat without asking again, except money.” Approve one client email, and later emails in that thread may simply go. Yet the same page says Approve “carries out that step, and only that step”. Until the vendor clarifies this, start a new chat for each sensitive batch and read the first result closely.
Where the StellarFirm Documents Disagree
Seven Conflicts We Found
A fast-moving product will always have stale pages, so none of this proves bad faith. It does matter when the disputed topic is who controls an autonomous system. We compared the live pages side by side, and the table quotes the conflict in each case.
| Topic | One page says | Another page says |
|---|---|---|
| Opening a pull request | FAQ: opens one “after your Approve” | Approvals table: does not ask |
| Merging | House rules: merges “always ask today” | Same page: auto merge on green, “no extra ask” |
| Scope of one Approve | “That step, and only that step” | Same kind of step, same chat, no new ask |
| Prices | llms.txt: “does not publish prices”; homepage: “Plans coming soon” | Pricing page: €19 and €59 plans |
| Phone apps | FAQ: “also available” | Homepage: “Not in the stores yet” |
| Developer access | Pricing: MCP and API “coming to Ultra only” | llms.txt: public MCP server with signed-in tools |
| Sales model | Astro Code site: “Buy once” | Product site: plans and credits |
Why the Merge Conflict Matters Most
The merge row is the one to settle before you connect production code. The house rules page offers auto merge on green, plus a separate rule to ship automatically after a merge. Together those read as a path from an AI patch to live code with no person involved, while the same page says merges always ask.
Check which behaviour applies on your account by testing it against a throwaway repository. Our report on an AI coding agent inside a secure Pentagon system shows why review steps around agents deserve this kind of scepticism.
Smaller Counting Slips
Some differences are merely untidy. The docs describe “fifteen” assistants while the homepage lists sixteen roles, and the starter page says “twenty” while its index lists 23. These are harmless on their own, but they suggest the pages are outrunning the people checking them.
StellarFirm Pricing: Plans, Tiers and the Launch Offer
What Each Plan Costs
The pricing page lists Hobby at no fee, Plus from €19 and Ultra from €59 a month, plus VAT. Business is custom. New accounts get 2,500 free credits and a 14-day trial with every model and Ultra’s limits, with no card needed. After the trial, Hobby is limited to open models such as GLM, Qwen, Mistral and DeepSeek.
Through November 2026, monthly plans are 50% off the first month and yearly plans 20% off the first year. Paying yearly already saves 20%, so a 1,000-credit Plus year drops from €228 to €145.92 in year one. Included credits do not roll over, while top-up credits stay until you use them.
The Plan Buys Access, Not Cheaper Credits
Top-ups cost a flat €0.01 per credit. Comparing that with each tier, a plan is largely a fee plus credits at about the top-up price. Plus only beats equal top-ups from 20,000 credits a month, and then by €1. Ultra costs exactly €40 more than Plus at every level we extracted from the page data.
| Credits a month | Plus | Ultra | Same credits as top-ups |
|---|---|---|---|
| 1,000 | €19 | €59 | €10 |
| 5,000 | €59 | €99 | €50 |
| 10,000 | €104 | €144 | €100 |
| 20,000 | €199 | €239 | €200 |
| 50,000 | €459 | €499 | €500 |
| 100,000 | €909 | €949 | €1,000 |
Dividing each Plus price by its credits gives the effective rate. Small tiers carry most of the plan fee, and the rate only falls below one cent at the top of the range.
A Possible Cheaper Tier
The page data also holds a zero-credit option at €9 a month, or €90 a year. That would fit the “platform fee” that the vendor’s llms.txt mentions for bringing your own models, but the visible pricing copy does not describe it. Ask the vendor about it before you budget around it.
What a StellarFirm Credit Really Buys
Our Markup Calculation
The FAQ says each model is “priced from what its maker charges”. That lets us test the margin, because the credit rate card lists credits per million tokens and the makers publish dollars. Dividing one by the other gives a remarkably steady ratio. This is our arithmetic, not a vendor figure, and it uses the ECB rate of 30 October 2026, where $1 is about €0.895.
| Model | Maker list, in/out per M | Credits, in/out per M | Credits per $1 |
|---|---|---|---|
| Claude Opus 5.5 | $4 / $20 | 572 / 2,858 | 143 |
| Claude Sonnet 5.5 | $2 / $10 | 286 / 1,429 | 143 |
| Claude Haiku 5.5 | $0.10 / $0.50 | 14.29 / 71.45 | 143 |
| Grok 4.7 | $2 / $6 | 286 / 857 | 143 |
| GPT-6.1 Sol | $2 / $10 | 314 / 1,572 | 157 |
| GPT-6 Luna | $0.10 / $0.50 | 15.72 / 78.6 | 157 |
So $1 of Claude or Grok list price costs 143 credits, or €1.43 at the top-up rate, against €0.895 if you paid the maker. GPT models cost 157 credits, or €1.57. The ratio of 142.9 also equals 100 divided by 0.7, which would fit a 30% margin if a credit were treated as one US cent, though that is our inference. Selling in euros then adds more.
A premium is not improper. Anyone reselling AI models carries payment costs, free credits and routing, and the vendor discloses every input needed to do this sum. The fix is also built in, because your own keys never use credits. Take $100 of Sonnet at list price. Through credits on Plus that is 14,300 credits, or about €147. With your own key it is about €89.51 plus the €19 plan, or €108.51, and you still hold 1,000 credits.
What One Ask Costs
The vendor’s example is 10,000 words read and 1,000 written, at its figure of 750,000 words per million tokens. Applying the rate card gives the bars below. Its own rounded figures are 0.5, 4 and 20 credits for unnamed small, balanced and frontier models, so ours run higher for Sonnet and Opus.
The Meter on the Computer
Model calls are not the only charge. An assistant’s cloud computer costs 0.4 credits a minute, or 24 an hour, while it runs. The 1,000 credits in entry-level Plus therefore cover about 41.7 computer hours and no model calls at all. One assistant working 8 hours a day for 22 days uses 4,224 credits, or €42.24, before it reads a word.
An always-on machine for 720 hours would use 17,280 credits, or €172.80, although idle workspaces go to sleep. Running assistants on your own computer through the desktop app avoids the charge. Search tools add 0.4 to 3 credits a call after 50 free calls a month per tool.
StellarFirm Data, Privacy and Security
Who Sees Your Work
The privacy policy is long and specific, which is a good sign. It also names a wide group of processors, and it says plainly that “this policy does not name” the model providers, which may process text outside the European Economic Area under standard contractual clauses.
| Data flow | What the policy says |
|---|---|
| Conversation text, cloud assistants | Recorded in PostHog and LangSmith, keys removed, can be switched off |
| Retention of that text | 30 days in PostHog; LangSmith period not stated |
| Assistants on your own computer | Text never sent to those records |
| Help chat, Diggy | Runs on Dify, may be processed outside the EEA |
| Sign-up events | Server may tell Meta, TikTok and Google Ads, never your email |
| Business visitors | Leadsy identifies companies, only after cookies are accepted |
Three Settings to Check on Day One
Go to Settings, then Privacy, and decide on Share model call analytics and Include conversation text. Both can be switched off, though the usage record for metering stays. Since the policy describes recording and then offers a switch, assume it is on until you see otherwise.
The word “train” does not appear anywhere in the privacy policy. It neither promises nor rules out training on your content, and it says records help “keep answers good”. Compare dikaio.ai, an earlier product the founders helped build, which states that data is never used to train models. Ask for that promise in writing before pasting client material into it.
Security and Compliance Claims
The security page describes separate workspaces, encrypted tokens that are never shown back, and a network allowlist. It offers no certification, audit report or bug bounty, so a cybersecurity reviewer has little independent evidence to go on. Business promises “help with” EU AI Act, HIPAA and ISO 27001 reviews, which is different from holding those certificates.
Europe needs particular care. Business lists EU data residency, but the European page says the service “will run in the EU” and that it “does not say the work is already done”. The same page says a data processing agreement will be offered. If you are subject to GDPR, that is a blocker until you hold a signed agreement.
Who Is Behind StellarFirm
Astro Code, a Small Greek Agency
The terms name Astro Code L.P. of Athens, with Greek law and the Athens courts governing disputes, while preserving mandatory rights to claim where you live. Users must be 16 or older. Liability for indirect loss is excluded where the law allows, and the service is provided “as available” with no promise of any feature or date.
The agency’s own site names founders Vasilis Drosatos and Nikos Benakis among a team of six, and calls this its first product. It also still advertises “Buy once. Bring your own API keys”, which no longer matches the credit model. That suggests the business model changed recently and quickly.
Checking the Track Record
The homepage cites past work with Dikaio.ai, Legitima.ai and Uizard. The agency site gives more detail, and that detail reframes the claims somewhat. We could not independently confirm its figures, so read them as the company’s own account.
| Claim | Agency’s account | What we could check |
|---|---|---|
| Dikaio.ai assistant | “Founding members”, partnered with Decode.gr, €20K MRR | Dikaio.ai’s site names Decode GR as its company |
| Legitima.ai assistant | Same engagement as Dikaio.ai | Live site now sells AI vendor compliance scans |
| Uizard acquisition | Ran growth team, 300K to 3M users | Miro did [buy Uizard in 2024](https://siliconcanals.com/uizard-acquired-by-miro/); role unconfirmed |
A two-founder agency launching this much surface in a month is impressive, and it also raises questions of durability. Our look at the Ollie AI assistant shutdown shows what customers lose when a young assistant closes. Keep your work in your own repositories and drives so that leaving is easy.
StellarFirm vs Lindy, Sintra and Paperclip
Rivals Side by Side
Plenty of products sell a team of digital workers. These three show the range: a credit-based rival, a role-based subscription and an open source manager. Credits differ between vendors, so entry prices below are not like for like, and we tested none of them.
| Factor | StellarFirm | Lindy | Sintra | Paperclip |
|---|---|---|---|---|
| Entry price | Free, or €19 + VAT | Free trial, then $29.99 | $97, on sale at $48.50 | Free local install |
| Included usage | 1,000 credits | 3,000 credits | 250 credits | Bring your own agents |
| Starting allowance | 2,500 credits, 14 days | $50 credits, 7 days | 14-day money back | Not applicable |
| Multi-user | Coming soon | Credits pooled, billed per user | Not compared | Hosted version on waitlist |
| Maturity signal | Domain from October 2026 | Shipping updates, enterprise tier | Twelve named helpers | MIT, 109,000 GitHub stars |
Where It Stands Out
Three things distinguish it on paper. The approval system is more thoroughly documented than most, and spending has a hard stop. It offers an unusually broad model list, with your own keys or your own computer exempt from credits. And the Coder’s first-week guide on its blog gives practical advice, including starting with a narrow token.
Where Others Lead
Lindy advertises 1,500 integrations, HIPAA support on its enterprise tier and audit logs. Paperclip lets you inspect the code and run locally, which a proprietary product does not allow. Sintra is simpler to budget for non-technical buyers. Any of these has more public history to examine.
Should You Use StellarFirm?
Who It Suits
Solo technical founders are the natural fit. If you can read a diff, have a small repository to experiment on and hold your own Claude or OpenAI key, the bring-your-own route removes the markup and the human check still has teeth. Small agencies might try it on low-stakes internal work once coworker accounts arrive.
Who Should Wait
Wait if you handle regulated or client-confidential data, need a signed data processing agreement, or need several people in one account. Anyone unable to judge the output should also hold back. A legal draft or a set of accounts is only safe in the hands of someone who could spot the error, whatever the approval screen shows.
A Careful First Week
The 14-day trial is long enough to test the claims that matter, and it costs nothing. Run these checks on a test repository and a test mailbox before you connect anything real.
- Create a fine-grained token limited to one throwaway repository.
- Set a low budget per company and per assistant, then confirm it pauses work.
- Switch off conversation text in Settings if you plan to use anything sensitive.
- Approve one email in a chat, then ask for a second, and see whether it asks again.
- Try auto merge on green on the test repository, and record whether a merge request appears.
- Compare the credits on three real tasks with your own key’s cost for the same work.
- Email the vendor about model training, the DPA and the merge rule, and keep the replies.
If the answers and behaviour match the strictest reading of the documents, it may well earn a place in a small stack. If they do not, you will have found out for free.
StellarFirm FAQs
Is StellarFirm Free?
Partly. Hobby has no plan fee, and new accounts get 2,500 credits and a 14-day trial without a card. After that, Hobby runs open models only and you pay for credits through top-ups at €0.01 each, plus VAT.
How Much Does StellarFirm Cost?
Plus starts at €19 a month and Ultra at €59, both plus VAT and both including 1,000 credits at that level. Prices rise with the credits you choose, up to €909 and €949 for 100,000 credits. Business pricing is custom.
Does StellarFirm Send Anything Without Approval?
Mail, messages and invites ask first, and real spending always asks. Pushing a branch, opening a pull request or issue and saving a new file do not ask. Your request text can reach a model provider before any approval, and one approval may cover the same kind of step later in that chat.
Can I Use My Own API Keys With StellarFirm?
Yes. Calls on your own key are billed by the provider and use no credits. Your own Claude, GPT, Gemini and Grok keys need the trial, Plus or Ultra. Other supported keys work on every plan, and models you run on your own computer are never charged.
Is StellarFirm Safe for Client Data?
We would not assume so yet. Conversation text from cloud assistants is recorded for debugging unless you turn it off, model providers go unnamed, the policy is silent on training, and EU data residency and a DPA are described as commitments. Get written answers first.
Who Owns StellarFirm?
Astro Code L.P., based in Athens, Greece. Its site names Vasilis Drosatos and Nikos Benakis as founders. The company describes it as its first product, and it runs as proprietary software rather than open source.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
