FAQ

Questions about Perch, answered.

What Perch is, how it verifies its work, and how the Web, Desktop, and CLI surfaces differ. If your question is not here, the docs go deeper, and you can always talk to us.

Getting started

What is Perch AI?
Perch AI is an AI agent you hand real work to. It reads the files you point it at, runs the commands the job needs, splits big jobs across a team of agents and checks the result before it answers. It works in the browser, on your desktop, in your terminal and in the cloud.
Is Perch free?
Starter is free with no credit card required. After sign-in you choose Starter, Pro or Ultra. Starter includes 20,000 PT per calendar month, which is up to $20 worth of included hosted usage on the Starter Roost pool. Pro is $10/mo during Early Access, half the $20 list price, and includes the premium lineup plus up to $150/mo of included usage, up to $75 of it on models you pick by hand. Ultra is $30/mo, with no 5-hour limit and three times Pro's room: up to $450/mo, up to $225 of it on models you pick by hand. The same account works across Perch AI Web, the desktop app, and the CLI.
Do I need an account to use Perch?
Yes. Create a free account at chat.perchai.app. On first sign-in you pick Starter, Pro or Ultra. There is no invite code and no waitlist.
How do I get started with Perch?
The fastest path is Perch AI Web at chat.perchai.app, which runs in the browser. For local folder access and agentic work, install the desktop app from perchai.app/desktop, or install the CLI with the command npm install -g perchai-cli. Sign in once and you are working in about a minute.

Products and platforms

What is the difference between Perch AI Web, Desktop, and the CLI?
Perch AI Web is a browser chat for grounded, cited answers, document questions, and web search. Perch Terminal Desktop and the Perch CLI add local file access, shell commands, sandboxed code execution, browser delivery, and multi-step agentic work. All three surfaces share one account, the same memory, and the same citation engine.
Is there a Perch mobile app?
Not in the App Store or Google Play. Perch runs in your phone's browser at chat.perchai.app, and you can add it to your Home Screen so it opens like an app, with its own icon. Every plan can chat from the phone. With Pro, in beta, you can also start cloud agents, send messages to your computer with remote control, and get a notification when a run finishes or needs your approval. The Perch on your phone guide in the docs has the steps for iPhone and Android.
Which operating systems does the Perch desktop app support?
The Perch desktop app runs on macOS (Apple silicon and Intel), Windows and Linux (x64, beta). You can download it from perchai.app/desktop, sign in with your Perch account, choose a folder, and start your first task.
What do I need to run the Perch CLI?
The Perch CLI installs with the command npm install -g perchai-cli and requires Node 18 or newer. After installing, run perch login once to sign in, then run perch inside any folder to work from the files there.
Can Perch work with apps like Gmail, Google Docs, and Calendar?
Yes, on Perch Terminal Desktop and the CLI. Perch works inside the browser sessions you are already signed into, so it can create Google Docs, draft and send Gmail, and add calendar events, and it can work in tools such as Outlook, Word, OneDrive, Slack, and Notion. Every action leaves a receipt you can review.

How Perch works

How does Perch verify citations?
Every factual claim Perch produces points back to the material it actually read. Claims are checked against their sources, and any line that cannot be supported is flagged or removed before it reaches your final draft. You can open each reference and confirm it yourself.
What are Saffron and Quill?
Saffron and Quill are two voices for the same Perch. They share one workspace, one memory, and exactly the same tools, so either one will audit a folder, reconcile a ledger, draft a memo, or check citations. The only difference is how the replies read: Saffron is direct and dry and leads with the finding, Quill is warmer and more literate on the page. Switch whenever you want a different read.
What is Flock?
Flock is Perch's deepest mode for hard tasks. Perch plans a custom team of agents for the task, runs them together and in parallel where the work allows, and has checker agents review the output before it reaches you. Unsupported claims are revised and rechecked, and the verdict travels with the result.
What is deep research?
Deep research is a slash command on Perch Terminal Desktop and the CLI. You ask a question with /deepresearch, and Perch plans it into sub-questions, searches your indexed files and the web across several rounds, then returns a cited report with a claim ledger that marks what is backed by your files, what is web only, and what is unsupported. /research is the same command. Details are in the slash commands docs.
What files and documents can Perch read?
On the desktop app and the CLI, Perch reads the files inside the folder you choose, and that folder sets the boundary for what it can access. On Perch AI Web, you can upload documents such as PDFs, spreadsheets, and images, then ask questions across them.

Plans and pricing

What is the difference between Starter and Pro?
Starter is free hosted Perch with a rotating list of frontier models, subject to availability, and 20,000 PT per month. You can pin any current Starter model by name on Desktop and the CLI. Pro adds Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Luna, Claude Sonnet 5.5 and Claude Haiku 5.5, fair-use Roost routing, the premium manual registry, Flock, and financial suites for $10/mo during Early Access, half the $20 list price. Both plans can also run on your own API key, which is not metered against your included usage.
What is Ultra?
Ultra is the plan for heavy daily use, at $30/mo. It includes everything in Pro, with no 5-hour limit, three times Pro's weekly limits on both usage lanes and three times its billing-period allowance for manual model selection, and two models only Ultra can pick: Claude Fable 5.2 and GPT-6 Astra. Claude Fable 5.2's host keeps prompts and replies for up to 30 days and may have people review them for safety. The Ultra plan guide in the docs lists the exact limits.
What is DeepSeek V4.1 Flash Unfiltered?
DeepSeek V4.1 Flash Unfiltered is the same model with most refusals removed, at nearly the same speed. It is for approved accounts on the Unlimited DeepSeek pass. To apply, open the model picker and click the Unfiltered card in the Unlimited group, under the pass card. A person reviews every application, and pass holders are reviewed first. Some requests are still screened. Breaking a rule removes Unfiltered access, and illegal use closes the account. The Unfiltered guide in the docs has the rules word for word.
Will the $10 Early Access price go up?
Early Access pricing for Pro ($10/mo) is available for a limited time, and we'll give notice before it closes to new subscribers. If you subscribe during Early Access, that $10/mo rate continues on each renewal for as long as your subscription stays active, and we change the price you pay only with advance notice. After Early Access ends, new subscribers pay the $20/mo list price.
How much can I use on Starter?
Starter includes 20,000 PT per calendar month, which is up to $20 worth of included hosted usage on the Starter Roost pool. Usage resets each calendar month in UTC.
Which models are included on each plan?
Starter's free lineup includes Laguna S 2.1, DeepSeek V4.1 Flash, and GLM 5.3 Flash, plus Qwen 3.8 Max until November 20, 2026. Usage still counts toward Starter's included allowance. GPT-6.1 Sol is Pro only. Pro includes every current Starter model and the premium registry except Claude Fable 5.2 and GPT-6 Astra, led by Gemini 4 Hydrogen, Gemini 4 Carbon and Gemini 4 Argon. Ultra adds those two. See the Models guide for the current lineup.
How do I upgrade to Pro?
Pick Pro on the plan chooser after sign-in, or open Settings → Billing in Perch AI Web and start checkout. Pro is billed through Stripe. You can manage or cancel your subscription from the same Billing panel.
Does Starter or Pro include direct API access?
No. Starter, Pro and Ultra include Perch-hosted model usage through Perch AI Web, Desktop, and CLI. Perch's internal application endpoints are not a public API and direct calls may be blocked or rate limited. Repeated attempts may result in account suspension. Contact us about an integration plan if you need direct API access. Bring-your-own-key usage inside Perch remains available on every plan.

Models

What model does Perch use?
Perch runs on Roost, its own model layer. Roost routes each task to a capable model automatically, so you choose a tier rather than a specific model and are never locked to one provider. On Starter, hosted inference uses a rotating list of frontier models, subject to availability. Pro unlocks Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Luna, Claude Sonnet 5.5 and Claude Haiku 5.5, the premium registry, and manual pins. Under the hood Roost runs leading models across labs, and the lineup keeps moving as stronger models ship.
Which models are available on Perch?
Roost draws on a rotating set of leading models and picks the right one for each task, so the exact lineup changes as the field advances. Starter has a rotating list of frontier models, subject to availability. Pro adds Gemini 4 Hydrogen, Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, GPT-6 Luna, Claude Sonnet 5.5 and Claude Haiku 5.5, and manual pins across every Pro model, including the current Starter pool. Ultra adds Claude Fable 5.2 and GPT-6 Astra. The Models guide lists current availability and rates.
Can I use images with a text-only model?
Yes. Some of the strongest models are text-only and cannot open an image at all, so Perch reads the image for them. Share a screenshot, a scanned invoice, or a photo with any model, and Perch describes what is in it and passes that to the model doing your work. Scanned PDFs are read the same way when they have no text layer. This works on every plan including the free Starter tier, there is nothing to turn on, and the reading step does not count against your included monthly usage.
What are the Roost tiers and context windows?
Roost offers Standard and Pro tiers, each with a Max variant. Standard and Pro use a 200K-token context window. The Max variants, Standard Max and Pro Max, extend context to 1M tokens for long documents and large codebases. Tier selection is available in the Perch Terminal composer.

Privacy and trust

Does Perch train on my data, and who can see my files?
It depends on the plan and inference source at the time of the interaction. Pro, Ultra and Enterprise activity, including complimentary Pro or Ultra access, is excluded from training and dataset licensing. Bring-your-own-key and self-hosted usage is excluded on every plan. Eligible Perch-funded Starter activity may contribute to Derived Training Data for AI training, evaluation, and benchmarking after aggressive deidentification and transformation. Perch never licenses raw User Content, personal information, email or browser content, credentials, or information that remains reasonably linkable to its source. To generate a response, the material needed for that request is sent to Perch's inference providers, which process it to provide inference and are not permitted by Perch to train on it or share it with underlying model vendors. Claude Fable 5.2 on Ultra adds one condition: its host keeps prompts and replies for up to 30 days and may have people review them for safety. Full detail is in the Privacy Policy.
Can I turn off the Starter Data Program?
Yes. The Starter Data Program is enabled by default after a versioned notice is shown, and you can turn it off at any time in Settings → Privacy. Turning it off stops future eligible collection and use and, where feasible, removes records that are still raw and have not been exported. It does not affect operational processing needed to provide and secure Perch. Data that is no longer reasonably linkable to your account may not be removable from a delivered dataset or trained model. The program applies prospectively from its September 13, 2026 effective date; historical activity is not included without separate explicit authorization.
Why should I trust Perch's answers?
Perch is built around verifiable output rather than confident guessing. On the web, each factual claim is tied to a cited source you can open. In the Terminal, figures are traced to tool results and marked as unsourced when they cannot be confirmed. The aim is work you can stand behind and check, not text you have to take on faith.
How does Perch handle my files and data?
On the desktop app and the CLI, Perch works within the folders you approve, and that folder boundary defines what it can read and change. For browser tasks, Perch acts through the sessions you are already signed into rather than asking you to hand over separate credentials. Every action it takes leaves a receipt you can review.

Get started

The fastest way to understand Perch is to try one real task.