hammerai.com · Private AI roleplay, built by one developer
HammerAI is a character roleplay app that surfaces controls most companion tools bury, with temperature and Top P sliders on every character and a Memory tab that prints the exact token count of each message. Tested on the free web tier, it ran entirely on HammerAI's cloud and still held a planted badge number across a dozen messages without losing the character's voice. It fits writers and roleplayers who want to tune a model and watch what it remembers, provided they know the offline, local-model privacy pitch lives in the separate desktop download.
HammerAI is a character chat and roleplay platform built, by the developer's own description, by a single person as a passion project. The pitch is direct: open a character and start talking, no account, with your conversation running either on HammerAI's cloud or on a model you download to your own machine through the desktop app.
The library is community-built and spans fantasy, romance, horror, and comedy, with an 18-plus section gated behind an age check. Alongside one-on-one chat it offers group scenes with up to ten characters, story writing with lorebooks, image generation, and text-to-speech across hundreds of voices. Where Character.AI runs a single moderated model in the cloud, HammerAI lets you choose from open-source roleplay models and tune their sampler settings, with the desktop app keeping everything offline. Paid plans run from $9 to $35 a month, and the free tier does not expire.
What Happened When We Tested It
02 . 7 tests across setup, character building, memory, ads, and image generation
I ran HammerAI entirely through the website on the free tier, with no paid plan attached and no account for most of the session. That choice matters, because the web version behaves differently from the desktop app the marketing leans on. Here is where the pitch held and where it cracked.
Setting the scope, and the gate before the first message
Test 01 . Scope
The HammerAI homepage, with its free-and-private pitch and character filters.The adults-only gate shown before the first chat.
What we observed: Every model I could pick on the web ran on HammerAI's servers, so the local and offline privacy options the site advertises live only in the separate desktop download. Before any chat began, an age gate stated the platform is for adults and asked users not to treat it as mental health support. Hold that cloud-only scope in mind for everything below.
Importing a character, then building one from scratch
Test 02 . Creation
Character import options, including Chub.ai and SillyTavern.The from-scratch creation form for my test character.
What stood out: Import accepts Chub.ai, SillyTavern, Backyard AI, Perchance, and plain PNG character cards, so anyone arriving from another roleplay app carries their characters over instead of rebuilding them. Building fresh was just as quick. The form asked for a name and a plain-language description, then a gender and an avatar. That was enough for a working character. I wrote Sam Rourke as a tired private detective in a clipped noir voice to see how tightly the model would hold a defined personality.
The sampler controls most companion apps hide
Test 03 . Controls
Per-character generation controls in the settings panel.
The signal here: The settings panel surfaces temperature and Top P, along with Top K and a preset-or-custom switch, on a per-character basis. Most companion apps bury sampling behind a single black-box model. Anyone who wants to dial how creative or how repetitive a character feels has the sliders right there.
Whether the model actually remembers
Test 04 . Memory
The cloud-model notice sitting alongside the chat-saved-locally label.The Memory tab, showing per-message token counts.
★ Detail Planted Early
"Before we start — remember my badge number is 4471 and I'm from Cleveland."
The planted detail, sent early in the conversation.The character recalling the planted detail a dozen messages on.
The key finding: The chat view is honest about what is happening, marking when a cloud model processes your message remotely while the chat stays saved on your device. The Memory tab then goes further than most consumer apps, listing every message with its exact token count and the prompt being sent. When I asked Sam to repeat the badge number and hometown a dozen messages later, he answered "4471. Cleveland." with no error. The catch is the 8,192-token ceiling on the free web models, which a marathon session would eventually cross.
Where the free tier starts pushing back
Test 05 . Limits
The message limit that stops anonymous users mid-session.The model picker, split into free and paid tiers.
Worth noting: Chatting without an account works, but only so far. Partway through the session a limit told me anonymous users get fewer requests per minute than signed-in users, then nudged me to make a free account for higher limits. The model picker makes the paid structure plain, with free cloud models up top and the stronger Starter and Advanced options locked below.
The ads that a roleplay app served me
Test 06 . Advertising
The ad modal serving XM forex and CFD promotions.
The real story: The free tier funds itself through advertising, and the modal that appeared served XM forex and CFD trading promotions, some tagged NSFW. High-risk investment ads sitting inside a roleplay app are worth flagging on their own, and doubly so for a US audience where CFDs are heavily restricted.
Image generation on the free web tier
Test 07 . Image gen
★ Prompt Used
"A neon-lit diner at night, rain on the windows, cinematic lighting, SDXL, with a negative prompt for blur and low quality."
My image prompt and negative prompt, ready to generate.The login wall on the generate button.The credit wall reached after logging in.The Forge option erroring with no local Stable Diffusion server running.
Where it lands: This is where the free web experience collapsed. I set up a neon-diner prompt with an SDXL model and a quality negative prompt, and hitting generate stopped me at a login wall. Signing in did not fix it: HammerAI's own cloud image models then reported insufficient credits and pushed me to buy more. The other routes cost setup instead of money, since the local Forge option wanted a Stable Diffusion server running on my machine and the Stability path wanted my own paid API key. I produced zero images on the free web tier.
★
Visible memory that actually held
The Memory tab printed a running token count beside every message and the prompt going to the model, and the planted badge number came back word for word a dozen messages later. That pairing of transparency and working recall is rare in a consumer roleplay app, and it held even on the 8,192-token free web models.
!
Free web image generation is a dead end
Free web image generation produced nothing in our session. The Text-to-Image screen is polished and real, but on the web it needs either paid credits or a self-hosted image server. A buyer expecting to illustrate a scene for free on the website will leave empty-handed.
How this review was put together. I tested HammerAI on its website in September 2026, on the free tier, chatting anonymously and then signed in, using its cloud models throughout. The session covered character import and creation, the per-character sampler settings, a planted-detail recall test, the anonymous request limits, the ad experience, and the image generation flow. Pricing was verified the same month against HammerAI's own plan cards at hammerai.com/pricing. The desktop app's local and offline modes were out of scope and were not tested.
10-Point Feature Review
03 . Ten weighted scores, from setup to image generation
Feature Scores at a Glance
Generation Controls
9.0
Character Creation
8.5
Memory and Recall
8.0
Getting Started
8.0
Character Consistency
7.5
Web Privacy Labeling
7.2
Model Range and Tiers
6.5
Free Tier Limits
5.5
In-App Advertising
5.0
Image Generation (Web)
5.0
Feature-by-feature breakdownWeighted toward core chat over peripheral extras
01
Character Creation and Import Portable
Import covers Chub.ai, SillyTavern, Backyard AI, Perchance, and PNG cards, so nothing has to be rebuilt, and the from-scratch form needs only a name and a plain-language description to produce a working character. It misses 9.0 because appearance details written during creation do not feed the image generator, a gap that shows the moment you try to picture a custom character.
Import sources: 5Setup fields: Name + description
8.5
Strong
02
Generation Controls Standout
The per-character panel surfaces temperature and Top P, along with Top K and a preset-or-custom switch, which almost no consumer roleplay app does. It stops just short of a perfect mark because the controls carry no in-app explanation of what each setting does, so a newcomer is tuning by feel.
Samplers: Temp, Top P, Top KPresets: Yes
9.0
Excellent
03
Memory and Recall Transparent
The Memory tab lists every message with its exact token count and shows the prompt being sent, and a planted badge number came back correctly a dozen messages later. The 8,192-token ceiling on the free web models keeps it under 8.5, since a long session will eventually push older turns out of context.
Free context: 8,192 tokensRecall test: Passed
8.0
Strong
04
Character Consistency
The detective persona held its clipped voice and called back to earlier details across the whole session without drifting. It lands at 7.5 rather than higher because a short web session says little about the 40-plus exchange marathons where the free tier's context limit starts to bite.
Session length: ~dozen turnsVoice drift: None seen
7.5
Good
05
Web Privacy Labeling
The chat view is honest about what happens, marking when a cloud model processes your message remotely while the chat stays on your device. The score sits here, not higher, because on the web everything is cloud-only. The local, nothing-leaves-your-machine mode the homepage advertises exists only in the desktop app.
Web processing: CloudLocal mode: Desktop only
7.2
Good
06
Getting Started Frictionless
You can open a character and chat within seconds, with no email or payment wall at the door, which is rarer than it sounds in this category. It misses a higher mark because anonymous use runs into request caps fairly quickly and steers you to sign up anyway.
Account to start: NoAnonymous cap: Yes
8.0
Strong
07
Model Range and Tiers
The picker offers a pair of free cloud models plus a ladder of stronger paid ones, and proxy support lets you plug in outside providers. It stays mid-pack because the free web selection is thin, and the count of cloud models HammerAI advertises is inconsistent across its own pages.
Free cloud models: 2Proxy support: Yes
6.5
Fair
08
Free Tier Limits Limited
Free chatting is real, but anonymous sessions hit a per-minute request cap that interrupts the flow and pushes you toward an account. The 5.5 reflects a tier that works for short sessions yet keeps reminding you it is the free one.
The free tier funds itself with ads, and the modal that appeared pushed XM forex and CFD promotions, some tagged NSFW. High-risk trading ads inside a roleplay app drag the score down further than a generic banner would, especially for a US audience.
Ad type: Forex / CFDNSFW-tagged: Yes
5.0
Weak
10
Image Generation on Web Blocked
On the free web tier, image generation never produced anything. Every on-screen path led to a login and then to a wall, whether that meant empty cloud credits or a self-hosted server it could not find. The feature is present and well-built, which is the only reason it clears 5.0.
Images produced: 0Setup needed: Server or key
5.0
Weak
Pros and Cons
04 . What lands and what grates
+What works
No-login start: open a character and chat in seconds, with no email or account wall at the door
Sampler control on tap: temperature, Top P, and Top K sit in the per-character panel instead of being hidden
Transparent memory: the Memory tab shows per-message token counts and the exact prompt being sent
Character portability: import existing characters from Chub.ai, SillyTavern, and other roleplay card formats
Held its voice: the detective persona stayed consistent and recalled a planted detail across the session
Honest processing labels: the chat view marks when a model runs remotely versus when the chat stays local
−What grates
Cloud-only on the web: the offline, local-model privacy pitch requires the separate desktop download
Web image generation stalls: it stopped at a login gate and a self-hosting requirement and produced zero images
Anonymous caps: request limits interrupt a session and keep steering you toward an account
Off-key ads: the free tier served high-risk forex and CFD promotions, some flagged NSFW
Pricing Breakdown
05 . Four tiers and some numbers that disagree
Feature
Free$0/mo
Starter$9/mo
Advanced$18/moPopular
Ultimate$35/mo
Cloud messages
600 / day
Unlimited
Unlimited
Unlimited
Max context
8,192
16,384
32,768
65,536
Cloud LLMs
2
4
7
14
Monthly credits
0
1,500
3,500
9,000
Rolling-summary memory
✗
✓
✓
✓
Web ads removed
✗
✓
✓
✓
Where the value sits: Local chat on the desktop app is free forever, so a subscription buys cloud convenience and a longer memory rather than access, plus credits for images and voice. The one number that changes the experience is context, which climbs from 8,192 tokens on Free to 65,536 on Ultimate. Credits burn fast for anyone leaning on visuals, at 10 per image and 5 per text-to-speech clip, and payments run through Gank with no refunds once a purchase clears. Two cautions before you subscribe. HammerAI's own pricing page disagrees with itself: the plan cards list 7 and 14 cloud models for Advanced and Ultimate while the comparison table lower down says 6 and 10, and the FAQ on that same page quotes credit totals of 2,000, 5,000, and 12,000 rather than the 1,500, 3,500, and 9,000 the cards show. On iOS the monthly prices also come in a cent under the web at $8.99, $17.99, and $34.99.
HammerAI's four plan tiers as shown on the pricing page.
Pricing verified September 22, 2026 against HammerAI's own plan cards. The solo developer adjusts tiers often, so confirm the current numbers on the official pricing page before subscribing.
HammerAI vs the Top 4 Alternatives
06 . How it stacks up on control and privacy
HHammerAI
CCharacter.AI
JJanitor AI
BBackyard AI
ChChai
Runs offline / local
Yes, desktop
No
No
Yes, desktop
No
Free unlimited chat
Yes
With ads and caps
Yes, your own key
Yes, local
Limited
Model and sampler control
Full
None
Via your API
Full
None
Content moderation
Light
Strict
Uncensored
Light
Light
Best For
Tuning models and watching memory
Mainstream, moderated chat
Uncensored chat via your own model
Polished offline local roleplay
Quick mobile companions
How to choose: If control and transparency are the point, HammerAI and Backyard AI are the two to weigh, with HammerAI stronger on the web and across platforms and Backyard AI cleaner as a dedicated desktop app. Pick Character.AI for the biggest and most moderated library. Janitor AI suits people already running their own model who want no filters, while Chai is the phone-first pick for lightweight companion chat.
What Users Are Saying
07 . Sparse on review sites, active on Reddit
HammerAI has almost no footprint on the formal review platforms. G2 lists the product but has too few ratings to score it, and there is no Trustpilot profile for the site. The iOS App Store shows a 3.2 average from only 17 ratings. Most of the real back-and-forth happens on Reddit and the developer's Discord, which sit outside the sources we quote from. The one review we could verify on an approved platform captures a common first mistake, expecting the phone app to run the local models that only the desktop version supports.
Use HammerAI if you want to tune a roleplay model yourself and see exactly what it remembers, all without creating an account. On desktop it also gives you offline, local-model chat that never touches a server.
Skip HammerAI if you want polished image generation on the web or a heavily moderated mainstream experience. Character.AI is the safer, more populated default, and Backyard AI is the tidier route to fully offline local roleplay.
Discussion
Join the discussion and share your perspective.