Ask your agent about a PDF, a scan, a photo, a video or a recording. It reads the file, answers, and cites the page, second or spot behind each answer. At the end it writes you a report.
Each example is a folder with a question and its files. You can run any of them with the commands below.
market-streetA 1906 film from the front of a cable car on San Francisco's Market Street, four days before the earthquake. The agent logs every vehicle that crosses in front of the camera, with the second it passes and a box around it, then counts each kind.
Watch the run README
fair-robotA 1939 film of nearly an hour. When does a golden robot first appear, and where is it in the frame?
Watch the run README
document-checkA certified copy of a Delaware filing, and a test copy with every mark painted out. For each copy you get the marks it has, where they are, and a box on each.
Watch the run README
lake-meadThirteen Landsat images of Lake Mead, from 1984 to 2024. When did the water fall most, and where? The agent boxes each shoreline the water has left since 1984 and names the first image it is dry in.
README
video-chaptersNASA's 1969 Apollo 11 film, cut into chapters. Each chapter gets a title, a start and end time, a line on what happens, and a frame.
README
hear-and-findWhen is "The Eagle has landed" said during the descent, and when does the flag first show on screen? You get both times, the frame, and a box around the flag.
README
apollo-descentNASA's typed transcript, twelve photos and the descent film, turned into a minute-by-minute account of the landing. Each line cites its page, photo or film moment.
README
boulder-chaptersA 1930s government film about building Boulder Dam, cut into chapters. The film has no chapter cards, so the agent finds the breaks itself.
READMEA replay opens as a chat. On the left you see what the agent thought and did, with Jev's answers in pink. On the right is the report it wrote.
agentvision.seeask a vision model about pages, frames or part of a picture.vision.judgeask Jev, TypeSafe's model, for a yes or no, a choice, a score or a box.vision.transcribeturn speech into text, with the second each word is said. It works with fal, ElevenLabs or OpenAI.The read tool keeps its name and arguments. It now reads PDFs, Office files, spreadsheets, images and video, and it tells the model when it cannot read something.
pi-vision adds no new tool. The agent writes code, and five skills show it how.
lookrender pages and cut frames with bash, then ask about them.findfind when something is said in a long video, or the second a thing first appears, and draw a box around it.transcribewrite one timestamped transcript of a long recording and search it with rg.verifycheck a PDF for each mark on a checklist, yes or no, with a box on its page.reportend the run with an HTML page that answers the question and shows the evidence.git clone https://github.com/Parcha-ai/agentrun && cd agentrun
npm ci --ignore-scripts && npm run build
npm run demo:vision -- examples/document-check # scripted: no key, no network
npm run demo:vision -- examples/hear-and-find --live # real models, your keysA live run needs GEMINI_API_KEY (or another provider's key), TYPESAFE_API_KEY for Jev, and FAL_KEY for speech. The agent's code calls ffmpeg, ffprobe, pdftoppm and pdftotext.
The agent runs shell commands as your user, so run it in a container or a VM.
To ask your own question, make a folder with question.md and inputs.json, or pass --input.
@parcha/pi-visionthe extension and its five skills.@parcha/agentrun-jevJev, for vision.judge.packages/harnessthe host the examples run on. It comes with the repository: clone it and run npm ci.Node 22.19 or later, ESM only. Install with @beta. Apache-2.0.