About Perch
The AI you can actually cite.
Perch is a verifiable AI coworker for legal, finance, research, and coding. It reads your files, does the work, and shows you the source behind every claim. If a finding cannot be traced, it gets marked, not buried.
Why it exists
Most AI tools produce text you cannot defend.
Perch began with a specific problem: the people most exposed by a wrong answer, lawyers who can be sanctioned for a fabricated citation, analysts who sign off on the numbers, researchers who stake a claim on a source, were being handed AI that could not show its work. Perch is built for exactly those users, and every design decision starts from their standard: if you cannot check it, it does not ship.
What it is
One coworker, across the surfaces where the work happens.
Web, desktop, CLI
Perch runs in the browser at chat.perchai.app with no install, as a desktop app for macOS and Windows that reads a local folder, and as a terminal CLI that works from the folder you are already in.
Flock
Perch's most powerful mode plans a custom team of agents for a hard task, runs them together, and has a verifier catch and fix mistakes before it answers.
Financial forensics
Perch reads financial files and runs the actual checks: AP audit, three-way match, duplicate detection, GL reconciliation, and statement audit, computed deterministically and benchmarked on real-world datasets.
Writing Studio
Perch drafts memos, briefs, and research with a checked citation on every claim, grounds each line to a real source, and will not finish a draft that does not hold up.
Extensible through MCP
Perch connects to external data through the Model Context Protocol. It ships with CourtListener for legal research and Playwright for browser automation, and you can add any MCP server.
What we believe
The interface should respect the work.
Files first
The useful context is usually already on your machine. Perch starts with folders instead of asking you to paste fragments into a chat box.
Sources before sentences
The answer matters less than the trail behind it. Perch keeps claims close to the documents that support them, and marks what it cannot verify.
Compute, don't guess
For work that has a right answer, like reconciling invoices, Perch runs the checks as code rather than predicting the result. A language model is the wrong instrument for arithmetic.
Personas with taste
Quill and Saffron do the same work with the same tools. What changes is the voice, because reading a hundred findings a week in a register you like is not a small thing.
The product
Two voices. One place to work.
Quill · Warm
Warm, literate, present on the page. The same audits, drafts, and folder work, in a register that reads like a person wrote it.
Saffron · Direct
Blunt and economical. Leads with the finding, skips the warm-up, and gets dry when the data deserves it.
Common questions
What people ask about Perch.
What is Perch AI?
Perch AI is a verifiable AI coworker for legal, finance, research, and coding. It reads your files, does the work, and shows the source behind every claim. Anything it cannot trace to a source is marked, not buried.
Who is Perch for?
Perch is built for professionals who have to stand behind what they produce: lawyers citing case law, analysts reconciling invoices and statements, and researchers compiling evidence. The output has to be checkable or it is not useful.
What makes Perch different from a general AI assistant?
A general assistant gives you confident text with no trail. Perch keeps every claim next to its source, runs deterministic checks for financial work instead of guessing, plans multi-agent teams with a verifier for hard tasks through its Flock mode, and refuses to finish writing it cannot support.
Where does Perch run?
Three places: the web at chat.perchai.app with no install, a desktop app for macOS and Windows that works over a local folder, and a command-line tool that runs from the folder you are already in with background agentic work and citation verification.
Which AI models does Perch use?
Perch is not tied to one model. It selects among frontier and open-weight models by running each candidate through a standardized harness of realistic professional tasks with planted issues, scored on grounding, tool use, and honesty rather than leaderboard numbers.
How is Perch verifiable?
Every claim is kept close to the document that supports it, unverifiable findings are flagged rather than hidden, and for financial work the checks are computed deterministically as code so the result is reproducible and auditable, not a prediction.
How much does Perch cost?
Perch has three plans: Starter is free, Pro adds premium hosted models and higher included usage, and Enterprise is for teams. Every plan can also run on your own API key. Full details are on the pricing page.
Get started
Perch is early, verifiable, and shaped around real folders.
The best way to understand it is to try one real task: a memo with sources, an audit folder, a messy export, a draft you keep rewriting.