HammerAI Review: The Companion App That Runs Without the Internet

The reason is the design. What is HammerAI? A hybrid of local and cloud: the same character library works with a one-click local Ollama or llama.cpp setup, which is unlimited, offline and private, or with hosted models. Storing chats locally by default is what sets it apart.

Most companion apps treat your conversations as their inventory, while HammerAI chat history stays on your own drive by default. For an app category built on people's private evenings, that should be the norm, and it's almost nowhere else.

It's not for everyone. Local models need a decent machine, the free cloud models are small, and the new iOS app is rated 3.2/5 from 17 ratings.

If you want a polished companion that remembers your week and lives on your phone, look elsewhere. If you want a story workbench you control, this is the one I'd pick.

The flight

My best memory of those four months happened on a plane. I was on a 3-hour flight to a June wedding shoot in Seattle, there was no Wi-Fi on board, and I ran a frontier murder mystery off my laptop. The guy next to me paid $8 for in-flight internet to check email.

I'd like to say I didn't feel smug about it. I felt a bit smug about it, which says more about me than about him. He was paying for a connection to reach his inbox, and my murder mystery needed nothing but a laptop battery.

It also changed how I thought about the app for the rest of the summer.

What was checked in September

The facts on this page that are not my own story come from two fresh checks in late September 2026. The first, on 23 September, covered the web version on the free tier as a guest, plus the Linux desktop app. The second, on 24 September, covered the web version on the free tier again, this time logged in with a Google account.

That split matters for how you read the rest. My own months leaned on local models, with a spell on the paid Starter plan in month two, while the September checks looked at what a newcomer meets today: as a guest, with a fresh free account, and on Linux. Where my story and today's pages disagree, I say so and give both with their dates.

The moment it started generating, my GPU fans spun up like a leaf blower and Moose left the room. I took it as the first review of the evening, and it was not a good one. The app, to its credit, didn't care.

For a local AI setup, fifteen minutes is quick. Most guides for running models at home read like assembly instructions for a wardrobe, and this felt closer to installing a video player.

What struck me more was what didn't happen. There was no login for local use. I'd been fighting client portal passwords and two-factor prompts all day just to deliver a wedding film, so opening an app that just said 'hi' felt almost suspicious.

It's the feeling of a diner that doesn't ask for your phone number before it sells you pancakes. You keep waiting for the catch. In this case the catch turned out to be hardware and disk space, and that part lives in the friction section.

How the front door looks today

As of 24 September 2026 the site is live at hammerai.com. HammerAI runs as a web app in the browser. A desktop app is offered for Windows, macOS on Apple Silicon and Linux, where it comes as deb, rpm and flatpak packages. iOS and Android apps are offered too.

The listing gives English as its language. Per a third-party review, the Android app existed before the iOS app. So anyone who used HammerAI on an iPhone before mid-July 2026 was using the website, not a native app, which matters for one of my own stories further down.

Signing in, or not

An account is optional: guests can chat with the free cloud models on the web right away. There was no sign-up wall before chatting. The desktop app is promoted with the line "No configuration or login needed", and that line is accurate.

If you do want an account, the sign-in window offers Google, Discord, Apple, Email and Phone. Email sign-in has Sign Up and Login tabs and asks for an email and a password. The password field allows up to 100 characters, which is generous for anyone with a password manager and slightly funny for everyone else.

The Google route is the quick one. Signing in with Google worked on 24 September 2026, and that path shows no captcha. A new account gets a random auto-generated username, the kind of name a stock-footage site gives an uploaded file.

The profile shows the Google avatar, a banner and a bio. It has Characters, Stories, Followers, Following, Account and Payments tabs. So the account is less a login than a small public shop front, which is worth knowing before a Google photo ends up on it.

The Account tab lists linked accounts, which here means Google. It offers an optional email password, set through a link: the tab says "click to send an email to set the password". It also offers optional phone linking.

My advice is to skip the phone. Five sign-in routes are four more than a character chat app needs to know about you, and given how much of HammerAI works without any account at all, the lightest touch is the right one.

On desktop, the first launch opens a setup screen: theme, font, language, an optional persona, scrolling and an adult-content toggle. It's a short screen, and I'd rather make those choices up front than hunt for them later. The only one I'd slow down on is that last toggle, which the content section covers.

Popularity is not subtle either. Cora, the featured philosophy character, shows 1.4M chats. That number tells you the library has real traffic, not that any given bot is good; the busiest stall at a flea market is rarely the one with the best stock.

Beyond characters, a Stories section shows public user-written stories with ratings. It's a different kind of browsing, closer to a fan-fiction shelf than a chat list. I think the ratings do more work here than on characters, because a story can be judged before you commit an evening to it.

The desktop app browses the same shelves. It also shows the web cloud catalog and the same character library, fetched online. That's convenient, but it means browsing is the one part of the desktop app that needs the internet, even when the chat itself doesn't.

What the form lets you set

The form starts with the look. The avatar can be an image or a short silent MP4.

A character gallery holds 10 images on top of that. The form has tags and a creator's note. The creator's note is the part I'd never skip, because it's where you tell other people how the character is meant to be played.

Then comes the writing. The personality field holds 5,000 characters, which is room for a proper backstory rather than a dating-profile blurb. The form has scenario and example-dialogue fields.

A character can have a first message plus alternate first messages. Alternate openings are underrated. The same character can open a scene two different ways, and you pick the take you want, like choosing between two slates of the same shot.

For anyone who hasn't, an AI "Generate Character" button drafts a character. I'd use it for a first draft and rewrite the personality by hand. Generated characters tend to read like stock footage: technically fine, instantly forgettable, and used in a hundred other projects.

Stories get their own tool as well. A "Write Story" tool lets users write their own stories. On phones, the iOS listing says users can create characters from scratch or import them.

Who owns the edit button

The oddest design choice is how edit rights work. Instead of an account, character edit rights use a "secret key": on the web it's a URL the user must save, and on desktop it's stored locally.

I like the privacy logic, and I'd still save that URL in two places. The page makes clear it's on you to keep it, and nobody wants to lose the keys to a character they spent an evening writing.

Nothing was published during the September check. On the tools alone, I'd rank this creator near the top of what I know. It asks more of you than a one-box creator, and it gives more back.

The dice are my favourite part. I let a d20 decide whether my sheriff fell through a rotten mine-shaft plank. He did, on a 2.

No model would have dared to drop the lead that hard on its own, and that's exactly why the dice belong there. They take the decision out of the story's hands, the way a real table does. I didn't argue with the dice, and the story was better for the fall.

It's the continuity error of text models: the villain reads the last page of the script and decides to save everyone some time. Regenerate fixes the scene, but you can't regenerate a model into being smarter. For a mystery, that ceiling is the one that matters.

What the cloud models say today

The guest check on the web used Cora, a featured philosophy-assistant character. The default free cloud model is Mistral-Nemo Instruct 12B. Cloud replies as a guest took about 3 to 6 seconds.

That's fast enough that nobody sits waiting, which matters more in a story than people think: a slow reply breaks a scene the way a long buffer breaks a cut.

Asked for a philosopher for a beginner, Cora answered: "Have you considered Immanuel Kant? His 'Grounding for the Metaphysics of Morals' ... it's written in clear, concise language." Calling Kant easy for a beginner is a doubtful recommendation. It's like telling someone new to editing to start with a feature documentary.

After being told the user's name and a preference for short books, Cora suggested Montaigne's Essays. That's a much better answer, and it shows the model listens within a conversation even when its first instinct is off. Cora also stays in character, with tea-pouring flourishes between the book advice.

The same character, logged in

The logged-in chat on 24 September 2026 was with Cora again, on the same free cloud model, Mistral-Nemo Instruct 12B. Logged-in cloud replies took about 3 to 5 seconds, so an account buys no speed and loses none.

The opening was the odd part. Given a name, a trip to the Alps and a question about weekends in a first message, Cora opened with "smiles warmly Ah, welcome back" and the name. That "welcome back" was odd in a first conversation.

In editing terms, that's a continuity error in the first shot of the film: the character greets a stranger like a regular. It's harmless, and it's also the tell of a small model filling the gap with the most common scene it knows.

The next reply was better. Asked about fondue and cooking, Cora gave a philosophical riff on fondue as a communal experience and used the user's name. It's a philosophy character turning cheese into a seminar, which is either in character or a warning, depending on how you feel about seminars.

The cloud banner

A banner warns when a cloud model is in use: "You are using a Cloud Model. Your messages will be encrypted and run on a remote server." I like that a lot.

The local model, for contrast

In the desktop chat's LLM picker, a downloaded local model appears under "Local (Ollama)" above the cloud models. The local model shows Context 2,048, a Local LLM Context Size of Auto and an extra Repetition Penalty slider. Once it's picked, the chat header switches from the cloud icon to the local model name.

That's a small detail, but it tells you at a glance whether your words are leaving the house, and I'd like every hybrid app to copy it.

The tiny local model was less helpful.

On it, Cora recommended "A Dictionary of Philosophy" by Edward N. Zalta for a beginner, calling it "an excellent resource for beginners". That book does not exist as described; it is a hallucination from a tiny model.

Put that next to my confessing claim jumper and the lesson is the same. The smaller the model, the more confidently it improvises, and it improvises plot twists and book titles with the same straight face.

The toolbox

The chat tools are Undo, Stop, Regenerate, Continue, Impersonate, Image for an in-chat picture, Author's Note and Dice. Each message has Pin, Edit, Delete and Play audio buttons, the last one for text-to-speech. The chat side panel has Settings, Memory, Character, Edit and Image tabs.

Group chats hold up to 10 characters. I'd keep a group to three or four, because past that even a good game master loses track of who is talking, and a small model loses track sooner.

A reviewer found the writing on the best community characters strong for the price, with messages costing no credits. I agree with both halves, and I think the second causes the first. A message you can regenerate as often as you like, without watching a coin counter, invites you to direct the scene instead of settling for the first take.

Without them, the local model would have called the mine owner "the mine owner" forever. That's the difference between a character and a props list. On small models, I think lorebooks matter more than raw context size.

The pricing page lists lorebooks as "Unlimited dynamic memory". For once, the marketing line undersells the feature.

A reviewer found the free 8,192-token context made characters forget earlier details in long sessions, and that lorebooks helped. That's fair, and I'd make it the first rule for newcomers: anything that must never change goes in a lorebook, not in the chat.

The Memory tab

Free has rolling-summary memory, but free users can only write and edit the summary by hand. The Memory tab says so plainly: "On the free plan you can write and edit this memory yourself, but it won't update automatically. Upgrade to a paid plan to have the summary update automatically as you chat."

During the check the tab showed "Memory - Locked" with 0 tokens. The Memory tab also has message trimming and pinned messages. It lists all messages with token counts, 313 tokens for 8 messages in the check.

Most apps treat memory as a black box. Here you can open the box and count what's inside, which is the same reason I trust scopes more than my eyes when grading colour.

What it remembered in the check

In the guest memory check, Cora correctly repeated the user's name and the short-books preference: "Of course, [name]! You told me your name is [name], and you prefer short books." That was within the context window, which is the easy case.

The logged-in check went the same way. Cora correctly recalled the name, the hiking trip in the Alps and the new fondness for fondue, all three in one answer. Oddly, the model that greeted a stranger with "welcome back" had no trouble with the facts it had actually been given, which tells you short-term recall is not the weak point.

The hard case is a long story that outgrows the window. That's where the hand-written summary and the lorebooks do the work, and where the free plan asks you to be your own script supervisor.

I'd call HammerAI's free memory honest rather than generous. It tells you exactly what it can hold, it shows you the token count, and it leaves the rest to you.

For guests there is nothing to try. As a guest the "Generate Image" button stays disabled; it needs an account and credits. The HammerAI image generator dropdown shows "Anime" and a model named after HammerAI itself, with an adults-only badge.

An account doesn't change the answer. On a new free account, image generation fails with an insufficient-credits error that tells you to "purchase more credits or upgrade your plan". So the button works, and the wallet doesn't.

I'd have liked one free sample image for new users. A disabled button or an error teaches you nothing about whether the images are any good, and faces are where the trouble shows, since extra fingers and melting hands tend to turn up there first.

There is no live calling. No live voice-call feature is mentioned on any official page; only cloud text-to-speech is offered.

If you want to talk to a character on the phone, this isn't that app. If you want a narrator reading the last reply aloud while you make coffee, it is.

My take is that in a local-first app, the paid media are the least interesting part. The chat runs free on your own hardware, while images and voice run on credits on someone else's servers. That split is fair, but it means media is where HammerAI looks most like every other companion app, and least like itself.

Ads are not part of the deal on desktop. Free has no ads on desktop. No ads appeared in the web sessions, including the short guest test. No paywall popup appeared, and no message limit was hit during that short test.

Where free stops

Characters forget the small stuff at 8K. My sheriff's deputy forgot he'd quit whiskey twice in one evening. He'd sworn off it, and then he hadn't, and then he hadn't again.

It's the continuity error every editor dreads, the tie that changes colour between two camera angles, except here it happened twice before midnight. For casual chat that's forgivable. For a mystery, where the whole point is who said what and when, it's the thing that sends you to the pricing page.

That's also why I'd pair the free tier with lorebooks from day one. The context window forgets; a lorebook entry doesn't. The deputy's sobriety belongs in writing, not in the model's short-term memory.

I'd trust the pricing page, because it's the one with a number on it. And 600 is plenty either way. Anyone who chats past that in one day is already the kind of user who should be running a local model.

My verdict on free

The free tier is very generous: 600 cloud messages a day without an account, plus unlimited local chat. I don't know of a more generous free offer among the companion apps I've covered, and I don't say that lightly about a category that usually meters every message like a parking garage.

The honest limit is quality, not quantity. The two free cloud models are small, and the free memory needs your hands on it. If that trade suits you, free HammerAI is a complete product rather than a demo.

I'd tell anyone curious to spend a few weeks on free before even opening the pricing page.

At $9 the maths is short. It's about the price of a diner breakfast, once a month, for a cloud companion that stops forgetting things quite so quickly. For a month of evenings, that felt like fair money to me.

Advanced costs $18 a month, with a 32,768-token context. It has 7 cloud LLMs, adding Mistral Small 3 24B, Ling 2.6 Flash and Qwen3 30B A3B.

Ultimate costs $35 a month, with a 65,536-token context. It has 14 cloud LLMs, adding Llama 3.3 Euryale v2.3 70B, Cydonia 24B v4.1, Hermes 4 70B, WizardLM 2 8x22B, Rocinante 12B v1.1, Aion 2.0 (DeepSeek V3.2) and Aion 3.0 Mini (DeepSeek).

What the model picker shows

The web model picker groups cloud models by tier. Free has Mistral-Nemo Instruct 12B and Llama 3 8B Lunaris at 8,192 context, and Starter has Mythomax 13B and UnslopNemo v4.1 12B at 16,384. Advanced has Mistral Small 3 24B, Ling 2.6 Flash and Qwen3 30B A3B at 32,768, and Ultimate has the big models, including the two DeepSeek-based Aion models, at 65,536.

In the Ultimate group, Rocinante 12B v1.1 has a 32,768 context, the one odd entry out. The picker also offers "Cloud - Proxy". The in-app model list has 14 cloud models on Ultimate.

How billing works

Prices are in USD and billed monthly. There are no annual plans; Gank lists "Subscribe 1 Month" only.

Web payments go through Gank (ganknow.com/hammerai). Gank accepts credit cards and other payment methods. Mobile subscriptions use App Store or Google Play billing.

Upgrades are prorated. Downgrades apply at the end of the billing period. I like both rules; they're the billing version of a client paying for the revision they asked for, not the whole re-edit.

The account's Payments tab keeps the paper trail: it lists subscription periods and credit transactions. It also asks anyone who bought on Gank to "Enter your GankNow email to link your account".

That link step is the one I'd do on day one. A subscription paid under one email and an account signed in under another is how a paid plan ends up looking like a free one, and sorting that out by email is nobody's idea of an evening.

Two sets of model counts

The pricing page doesn't agree with itself.

On the pricing page the compare table lists 2/4/6/10 cloud LLMs, while the tier cards list 2/4/7/14. Gank's tier descriptions list older model names, such as Hermes 2 Pro 8B, Dolphin 3.0 and Magnum v4 72B.

In editing terms, that's several versions of the same timeline exported on different days. None of them is a lie, but you shouldn't pay for a specific model count, because the page itself can't agree on one. Pay for the context size and the models you can see in the picker.

Is the step up worth it?

For me, Starter was the whole answer. It's where the money makes a visible difference: the daily cap goes, the context doubles, and the phone gets a better brain.

Advanced and Ultimate buy bigger models, and the jump in context is real. For a desktop user with a good graphics card, though, the local models already cover most of what Advanced adds. I'd only pay for Ultimate if a phone or a weak laptop is where you chat and the small models have become the bottleneck.

My rule: $9 to feel the difference, $35 only if you're chasing the big models and know why.

That zero is the honest part of the page. Nothing is dangled in front of a new user, and nothing is handed out to hook them; the meter starts empty and stays that way until money goes in.

Plan credits

Starter includes 1,500 credits a month, enough for 150 images or 300 TTS generations. Advanced includes 3,500 credits a month, which is 350 images or 700 TTS. Ultimate includes 9,000 credits a month, which is 900 images or 1,800 TTS.

One image costs 10 credits. One TTS generation costs 5 credits. So a narrated reply is half the price of a picture, which is about the right ratio for how much each one adds to a story.

Plan credits refresh at the start of each billing period. Monthly plan credits do not roll over. Purchased credit packs never expire.

So plan credits are a use-it-or-lose-it allowance, like a gym class pass, and packs are the gift card that sits in a drawer until you need it. If you only want images now and then, the drawer is the better place for your money.

The Large pack costs $19.99 for 3,000 credits, or 300 images or 600 TTS. The Mega pack costs $49.99 for 8,500 credits, or 850 images or 1,700 TTS. The Ultra pack costs $99.99 for 20,000 credits, which buys 2,000 images or 4,000 TTS.

Doing the maths out loud: the Tiny pack's 20 images come to about ten cents each. The Ultra pack's 2,000 images come to about five cents each, which is a bulk discount only a very dedicated person would ever reach.

For most people the plan credits are the better deal. Starter's $9 buys unlimited cloud chat plus 1,500 credits, while the Medium pack alone costs $9.99 for 1,300. If you want images every month, subscribe; if you want a handful once, buy the Tiny pack and stop there.

The FAQ disagrees

The pricing FAQ gives monthly credits of 2,000 / 5,000 / 12,000, against 1,500 / 3,500 / 9,000 on the plan cards. That's roughly a third more in the FAQ for every tier.

No daily check-in was seen, and no referral program was seen either, logged in or not.

The paid side is just as quiet. No free trial was seen. No first-purchase discount was seen.

In most companion apps, this section is where the warnings about countdown timers that reset on every reload go. HammerAI has nothing to warn about. The price on the card is the price, every day, for everyone.

The missing free trial is less of a loss here than anywhere else, because the free tier is the trial. 600 cloud messages a day and unlimited local chat tell you more about an app than a three-day pass with a credit card on file ever could.

If you want to save money on HammerAI, the discount is structural. Run a local model, and the chat costs nothing.

Pay for Starter only in the months when you need the phone and the bigger context, then downgrade at the end of the billing period. That's the only coupon on offer, and I'd take it over a flashing banner any day.

What is switched on by default

HammerAI adult content is controlled by a switch in Settings > General, with Enabled and Disabled positions. That switch was enabled by default for a guest after the click-through. It also defaulted to Enabled in the desktop onboarding.

I don't love that default. A setting that changes what a stranger's bot can say to you should start off, and the user should switch it on as a choice. Turning it off takes a few seconds, so it's a small fix, but defaults are what most people keep.

Tools to narrow things down do exist. The home page has a "Safe" filter toggle. Users can set "Excluded Character Tags". Characters are rated Safe or Mature by their creator.

The ratings are only as good as the people applying them. The FAQ category list includes fetish categories by name. So the tag exclusions are the tool I'd set up first, before the library starts making suggestions.

The rules on paper

The content guidelines require every character to be an adult, and any character presented otherwise is removed. Characters marked Mature may explore adult themes within listed restrictions.

The Compliance page says six kinds of content are actively filtered and blocked: content involving anyone below the age of majority, self-harm encouragement, graphic violence, illegal activities, harassment and hate speech. Self-harm detection shows a crisis referral.

The footer tells a different story. It promises "raw, uncensored conversations ... no filters and no limits". That footer promise contradicts the Compliance page's claim of active filtering.

One page is marketing and one is policy; I'd read the policy as binding, while noting that the marketing is what a visitor sees first.

The front page

Several top characters in the home page's "Most liked" row carry settings and tags that conflict with the platform's own rules on the age of characters. Ages were not stated on those character cards. Those characters were not opened during the check.

The local catalog

So is HammerAI NSFW, free, unlimited and uncensored? On local models, largely yes. 81 of the 87 local models carry an adults-only badge, for example Gemma 4 Heretic, Cydonia, Omega Directive, Anubis 70B, GPT-OSS 120B and Qwen3.5 Abliterated 122B. Nearly all local catalog models are marketed as uncensored, abliterated or built for adult roleplay.

Some local model descriptions use phrases like "without ethical restrictions". Local chats are not moderated by the server; they never leave the device.

That's the trade in one line: the same privacy that makes local mode trustworthy means nobody is watching what runs there. For an adult who chose it, fine. For a shared household computer, it's a reason to think twice about where the desktop app goes.

It's the software version of checking whether a used car runs with the dealer out of the room. Privacy policies are words, and a chat that keeps replying with the cable lying on the desk is evidence.

Where chats live

Chats save to the browser's local storage by default, marked "Chat saved locally". The "Cloud Account" save location is greyed out for guests.

An account changes the options, not the default. Logged in, the chat header still said "Chat saved locally". Settings > Chat Save Location then allows "Cloud Account".

I like that order: the cloud is something you opt into, not something that happens to you the moment you sign in.

The account side keeps its own log. The Account tab lists active sessions with the OS, browser, IP and a 30-day expiry. That's a useful audit trail for anyone who signs in on a shared or borrowed machine, and it's also a reminder that the account, unlike the local chats, lives on a server.

That's the local story, and the cloud story has a footnote. Privacy policy section 12.1 says "HammerAI does not use user data to train AI models internally." Section 13 says cloud messages go to third-party providers, such as those reached through OpenRouter, which "may process and retain your messages ... including but not limited to AI model training".

The FAQ says cloud messages are "not stored or used for training", which reads as a statement about HammerAI itself. Put together: HammerAI doesn't train on your chats, but the providers behind the cloud models might. If that matters to you, the answer is local models, which is where my unplugged cable pointed anyway.

Who you're dealing with

The service is operated as "HammerAI"; no legal entity is named on the site. The App Store lists the seller as Sunspot LLC (Apps), and the app's company on record is Sunspot LLC.

The pricing page says HammerAI is "built by one developer as a passion project". The contact address is [email protected].

One developer explains a lot: the generous free tier, the pricing page that contradicts itself, and the manual updates on Linux. I'd rather deal with a passionate one-person shop than a faceless one, but it does mean fewer people to answer an email.

Payments go through a separate company. Gank is run by Gank Pte. Ltd., Singapore. Gank, the payment processor, collects and stores payment information for HammerAI.

Money, refunds and disputes

All purchases are final: "All purchases are final. GankNow does not support refunds".

The terms do not limit EU consumer rights, including the right to withdraw from distance contracts within 14 days. If you're in the EU, that line is worth a screenshot before you pay.

The terms are governed by Delaware law. Disputes go to binding AAA arbitration. The terms include class-action and jury-trial waivers.

Age rules and your characters

The minimum age is 18, or the age of majority or consent in the user's jurisdiction, whichever is greater. The age notice forbids using the app as a therapist or to discuss mental health. The terms have users confirm they do not find sexual material offensive or objectionable.

Published characters and stories grant HammerAI a "non-exclusive, royalty-free, perpetual, irrevocable" license. The terms note that purely AI-generated output may not be copyrightable, per US Copyright Office guidance. Rights to feedback are assigned to HammerAI.

In plain English: publish only what you're happy to give away for good. Private characters and local chats are yours; the moment something goes public, the licence goes with it.

Data processors and ads

Named data processors are Cloudflare, Sentry, Supabase, Vercel and self-hosted Umami analytics. The privacy policy says third-party advertising services, Adsterra and Growl, display ads on the website. It names those ad networks and says their use may count as a "sale" or "sharing" under the CCPA.

The privacy policy was last updated on March 29, 2026. The ad clause is the one I'd watch. The desktop app is the clean way in, and the website is where that permission lives.

A warning at the download button would have saved me a very specific kind of evening.

Local setup on Linux

Ollama is not bundled: on first use the Models page says "Before you can download an LLM, you need to download Ollama". The first Ollama download (v0.30.6) failed silently: the progress bar filled, nothing was installed, and the log said "Zip file does not exist". Retrying with v0.30.5 worked, at about 2.1 GB unpacked.

A progress bar that fills and delivers nothing is the worst kind of bug. It's the render that reaches the end and exports a black frame, and you only find out when you open the file.

Local generation was very slow in the check: about 1 token/s on a shared 2-CPU machine with no GPU, and CPU-only use is very slow. The Electron auto-updater is not supported on Linux, so Linux users update manually from GitHub.

Web sign-up, credits and the store links

The Sign In button is disabled for a few seconds after the page loads. The email sign-up form adds a Cloudflare Turnstile checkbox labelled "Verify you are human". Neither is a real obstacle for a normal visitor, just two small pauses at the front door that make the site feel slower than it is.

No explanation is shown for the disabled guest image generation. Logged-in free users get the "Insufficient credits" error only after clicking Generate, because they have 0 free credits. That's the order I'd flip: tell people the price before they write the prompt, not after, the way a good client brief puts the budget on page one.

Shaky answers and a messy pricing page

The free-tier cloud model gave shaky facts; it called Kant's Groundwork easy. The tiny local model invented a book. The pricing page contradicts itself on model counts and credit amounts, and the Gank tier texts are outdated.

None of these is a deal breaker alone. Together they read like a one-person workshop: excellent tools on the bench, and a floor nobody has had time to sweep.

If you're one of those, I'd install it tonight. If you're not, the reasons to skip are just as specific.

It's not for anyone who needs refunds, since all sales are final, or annual plans, or strong built-in content safety, with adult content on by default and an age gate that's only a click. It's not for people who want a polished companion with automatic long-term memory on the free tier, since the auto summary is paid only.

A reviewer found the many options, from local and cloud models to credits and lorebooks, confusing for new users despite the generous free tier. I think that's fair. HammerAI rewards people who enjoy a settings menu and quietly punishes people who don't.

Cancelling on the website won't touch a plan you bought in the app, and the reverse is true too. It's the most common way to pay twice for anything, and HammerAI is no exception.

The pricing page carries no clear sentence about auto-renewal. With no annual plans on offer, I'd assume a monthly renewal until told otherwise and set a calendar reminder a few days before it. The pause option is the underrated part: if you only want Starter for a busy month, pausing beats cancelling and resubscribing.

Deleting the account

Account deletion sits in the Settings > Account tab. That tab has a Sign Out button and a "Delete your account" option side by side, so read the label twice before you click.

Deleting the account removes the profile, cloud and private characters, and stories. Account deletion is irreversible.

Local chats stay on the device after account deletion. That cuts both ways. Your local stories survive a mistaken deletion, and they're also not erased by a deliberate one, so if you want them gone, delete them on the device too.

None of that takes long, and all of it is harder after the fact.

Then set the reminder. All sales are final, so the reminder is the only refund policy you really have, and I'd rather lose a minute to a calendar than a month to a forgotten renewal.

The Linux .deb package for that version is 105 MB, and the app is built on Electron.

Don't read "GitHub" as "open source". The GitHub repo hosts releases only, and it has 17 stars.

The desktop app is closed-source and built on Electron. It's a download shelf, which is fine, as long as you don't expect to read the code.

On phones, use the official store listings: the App Store and Google Play pages are both linked from the site. A different Play app called "HammerAi" (id ai.hammer, developer WhatsGood) is unrelated. Check the developer name before you install anything with a hammer in the icon.

Setting up local models

After install the app starts its own Ollama server on localhost. The Ollama tab lists Ollama releases (v0.30.6 and older) with one-click download, plus AMD ROCm packages. It can point to an external Ollama URL and set environment variables.

If you already run Ollama for other things, I'd point HammerAI at it rather than installing a second copy. Two servers fighting over one graphics card is not a relaxing evening.

The Ollama tab also sets the KV cache type (FP16, Q8_0 or Q4_0), Flash Attention and Num Parallel. "Test Ollama" and "See Ollama Logs" buttons help with troubleshooting. If a download finishes and nothing works, open the logs before trying anything else; they say more than the progress bar does.

Choosing a model

The LLMs tab has a catalog of 87 curated local models. Local model downloads range from 1.8 GB (Dolphin Phi 2.7B) to 81 GB. In the check, Dolphin Phi v2.6 2.7B (1.8 GB) downloaded in seconds.

"HuggingFace Download" fetches any GGUF model. The local model folder is configurable. Point that folder at a big external drive before the first download; it's the cheapest advice in this review, and it's the one that saves your main drive.

A tiny model will chat happily, but it will also invent books and confess to crimes it hasn't committed yet.

Before you go offline

Offline use needs a model downloaded beforehand. Download it at home, open one chat to make sure it loads, and only then trust it at cruising altitude. An offline app with no model on it is a very quiet app.

For the cloud side, a Proxy tab lets users bring their own API. There is no public API for HammerAI itself. So the proxy is a way in, not a way out: you can plug other models into HammerAI, but you can't plug HammerAI into your own tools.

Where to get help

The community lives on Discord, and HammerAI also runs an X account, a subreddit and a TikTok. Payments and support run through the Gank page.

For anything about a subscription, I'd start with Gank rather than Discord, because that's where the billing lives. For anything about a model that won't load, Discord and the logs will get you further than any support form.

Images and voice rate Weak. They are credit-metered extras with no calls, and a guest can't try a single image. Privacy and terms rate Good, because local storage by default is rare, but cloud messages pass through outside providers and the policy allows ads on the website.

Cancelling and billing are Mixed. Gank makes cancelling and pausing simple, but every sale is final and the pricing page never spells out renewal. Content settings rate Weak: adult content is on by default, the age gate is a single click, and the front page doesn't match the platform's own rules.

Where it wins is privacy, control and the free tier. Of the apps on this site, it's the one where the settings menu feels like it was written for the user rather than for the billing team. I don't know of another consumer app that lets you watch the token count, edit the prompt and pull the network cable in one evening.

If you want a HammerAI alternative that feels like a person, most of the table beats it. If you want an app that feels like a tool you own, very few do. I'd put it next to SillyTavern rather than next to the companion apps, and it's friendlier than SillyTavern by a wide margin.

More on this site: the best AI girlfriend apps, free AI chat apps, the best AI companion apps, AI roleplay chat apps, all AI companion app reviews.

It's not the smartest app in this category, and it's not the prettiest. The models are small, the pricing page argues with itself, and the front page needs a moderator with a free afternoon. I'd still take it over most of the polished apps, because the polish on those is often paint over a meter.

What stays with me is the quiet. A laptop, no network, and a story that belonged to nobody but the person telling it. That's a small thing to ask of software, and it's rare enough that I'd pay for it, even though I mostly don't have to.

The desktop app stays installed, and that says more than any score.