AI agents on the web, on your desktop and in your terminal
npm install -g perchai-cliStarter is free · Mac, Windows and Linux
Routines Beta for Pro
Write a task once and choose when it runs: on a schedule, or when your own code calls it. Perch does it in the cloud with your computer off, then tells you what it did.
Read the Routines guideArtifacts On every plan
Ask for a dashboard, memo or table and get one you can open beside the chat. Ask for a change and Perch saves a new version. Share any version with one link.
Read the Artifacts guideHand it the jobDesktop
Point Perch at a folder and describe the outcome. It plans the work, reads every file, runs the checks, and hands back a finished report.
100 files read93 findings$147,420 flagged
Real recordingRuns a teamDesktop
Perch splits a big job across helper agents that work in parallel, then checks their findings itself before it reports back.
3 helpers113 source filesGPT-6.1 Sol
Shows its workEvery finding
Each finding names the file, the row and the amount, so you can check it in seconds. This one is a $94,890.13 duplicate payment that skipped the three-way match.
INV-88421march_payment_run.csv, rows 8 to 9
Cites every claimWeb
Perch AI answers with numbered sources and marks anything it could not verify, so you know exactly what to trust.
5 sources1 claim marked unverified
GPT-6 Sol, GPT-6 Luna and GPT-6.1 Sol were all in Perch on launch day. Roost picks the right one for each task, or you pin one by name.
Start a job on the web, pick it up on your desktop, check it from your phone.
npm install -g perchai-cliWe test models on real work before they reach Perch, and we publish what we find.
ResearchJul 2, 20269 min read
Everyone asks which model is best for coding. We ran nine of them, closed and open, through the same build task and the same rubric. GPT-5.5-Codex was the most consistent, but an open-weight model won one of the two tasks outright. Here is the full leaderboard and how to read it.
Read the article →The strongest reasoning models are text-only. The standard fix is to swap in a multimodal model, which silently downgrades every turn of your agent, including the ones with no image in them. There is a better shape, and it has rules.
Aug 5, 2026 · 9 min readThe word "memory" covers three different things: the context window, retrieval over a corpus, and durable facts that persist across sessions. They fail differently, and most products build one and call it memory. Here is the full picture, and the four problems any system has to solve.
Aug 4, 2026 · 10 min readMuse Spark 1.1 launched claiming the agent and tool-use crown. Grok 4.5 launched claiming token efficiency. On document-heavy forensic work, one of those claims holds and the other inverts. The higher-scoring model was also the one that crashed.
Jul 19, 2026 · 8 min read
Frontier models on one plan, and the same Perch on every screen.
$10$20 per month
Early Access price, locked for as long as you stay subscribed.
Premium models
Need more room? Ultra is $30 a month, with no 5-hour limit.
Freeno card needed
Everything you need to put Perch on real work.
Bring your own API key on any plan. Compare plans and limits →
