Frisket

I wanted a spreadsheet that did AI things for investigative journalism. The NYT made one for itself, so I made one for the rest of us. Use Frisket to analyze zillions of documents, decades of audio, centuries of video, all in a friendly (???) spreadsheet format.

Quickstart

Use Frisket Desktop, Frisket Cloud, or Frisket Self-Hosted.

What Frisket can do

Frisket can analyze spreadsheets, PDFs, images, videos, probably a hundred other things. You drop content in, then run AI against it in a structured way. It's a new way to do projects like:

• Extract and cross-reference every company mentioned over thousands of legal documents
• Find every time datacenters are mentioned in a YouTube channel of city council meetings
• Split a video at scene changes to analyze popular streamers
• Download and translate TikTok videos to track misinformation
• Get an email or Slack message summarizing this week's new podcast eps
• Convert recipes into structured lists of ingredients and amounts (it isn't journalism but hey)
• Summarize, categorize, restructure, and clean documents
• Automatically perform cited research on the web

Those were all done the old-fashioned hard-work way, though. No one has done anything with Frisket yet, so you can be the first.

Extract similarly formatted documents

Open a document sheet's Extract view and draw a box around a label, then its value. Fields become columns. For repeated records, mark the first record and the remaining-records area; each record becomes a row in one combined result sheet. Preview a sample before extracting the collection, and click preview values to inspect their source regions. Save the template to reuse it.

Native PDF text works directly; scanned PDFs and images need OCR with positioned text first. Extraction itself uses no generative model or paid OCR. Expand value areas follows neighboring labels; Continue across pages is off by default. An empty value remains empty, and a document with no repeated records adds no rows.

Installation

Frisket Desktop

• Download for macOS — macOS 14 or newer, Apple Silicon
• Download for Windows — Windows 11, x64

Updates are manual: quit the app, install the new package, and reopen it. Your workspace and downloaded caches stay in the app profile.

Frisket Cloud

Open Frisket Cloud to request access.

Frisket Self-Hosted

Single-user

If you just want to run Frisket on your own computer, use solo mode. This requires Python 3.12 with SQLite 3.43 or newer and FTS5 support. Frisket Desktop and the Docker image include a suitable runtime. For a Python installation, check the linked SQLite version with python -c "import sqlite3; print(sqlite3.sqlite_version)". If it is older, use a recent Python distribution, such as one installed by uv python install 3.12.

pip install 'frisket-data[standard]'
frisket ./my-workspace

If you want more shiny extras, use pip install 'frisket-data[complete]'.

The package name is frisket-data, not frisket.

Team

If you're running Frisket on a server or want to support multiple users, go for team mode.

1. Download the server install bundle (frisket-server-<version>.tar.gz)
2. Extract it, then run sudo ./frisket-install from the extracted directory
3. The installer asks a few questions and then you should be good to go. It can keep Frisket behind an SSH tunnel or serve a public domain using Caddy for HTTPS.

Codex claims it requires Linux, Docker Engine, and Docker Compose. It installs Frisket under /srv/frisket.

Project Ask workflows

In Solo or Team, open Ask’s settings and enable Run actions to let it convert documents, extract fields, and analyze the results. Installed plugin actions use the same action system as built-ins.

Choose Ask before each action, Ask before overwriting, or Full access, and set a total budget for the conversation’s next research run. Leaving the budget blank uses your existing action approval limit. Leave Max turns blank for no turn limit. Work continues when you close the panel or browser, while the Frisket server and its worker remain running. A server restart interrupts the run; it does not automatically replay actions.

Use Settings → Skills to create, upload, edit, and enable an instruction-only SKILL.md. Team administrators manage the shared library. You can also ask Ask to save or revise a workflow as a skill when you have permission to manage the library. Ask creates a draft; Review the text opens the complete instructions in a full-size editor. Saving an enabled skill makes the revision available to your next message in the same conversation, without running the workflow by itself. A disabled skill remains unavailable to Ask until it is enabled in Settings.

Ordinary Ask turns can use enabled skills. Research settings can narrow that list for research runs. Each new turn takes a fresh snapshot of the enabled, selected instructions, while work already in progress keeps the revision it started with. A file starts with a name and description:

---
name: review-filings
description: Extract and compare facts from selected court filings.
---
At use time, inspect the selected sheet, columns, and row scope. Convert the
selected PDFs to Markdown, inspect a small sample, then extract the parties and
filing dates into new columns over the requested rows. Use the available filter
and analytics tools to count matching filings and cite the supporting records.
Discover the current action and its parameters before preparing it. Do not
overwrite existing columns or expand a selected row scope.

Skills guide the agent’s existing tools; they do not install or run Python or shell scripts, grant action access, or override the selected sheets and rows. Write reusable steps that resolve sheet, column, row, model, and action identifiers at use time. Do not put credentials or secrets in skills. You can also configure Exa or Tavily in provider settings for web search. Ask uses Exa first, then Tavily, when their keys are available.

Project Ask traces

Project Ask can send its agent traces to Braintrust or another OTLP backend. Set a standard OTLP HTTP endpoint and headers on the Frisket app process:

export OTEL_EXPORTER_OTLP_ENDPOINT=https://api.braintrust.dev/otel
export OTEL_EXPORTER_OTLP_HEADERS='Authorization=Bearer <API key>,x-bt-parent=project_id:<project id>'
export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf

The generic endpoint gets /v1/traces appended. If you use OTEL_EXPORTER_OTLP_TRACES_ENDPOINT, include the full trace path instead. OTEL_SDK_DISABLED=true or OTEL_TRACES_EXPORTER=none disables export. An unsupported protocol or malformed header also disables Ask tracing and logs a warning that names the invalid setting without logging its value.

Only the Project Ask chatbot is instrumented. Prompt, completion, and tool payloads are excluded by default; traces still include model, token, timing, and tool-name metadata. Set FRISKET_ASK_TRACE_CONTENT=true to include those text payloads for debugging after confirming the destination may receive project content. Binary media is always excluded. Frisket does not export actions, workers, metrics, or logs through this integration.

CAVEAT: The big, fancy parts

Frisket is split into a few parts, including a lightweight server and a heavier sidecar to optionally offload intensive work. As a result, different installs have slightly different features.

Local Docling needs no Docker setup. In Solo mode, choose Docling in the PDF-to-Markdown engine picker and click Download and set up. Frisket installs its CPU runtime in a private environment in your user cache, then starts the model server. The server source ships with Frisket; setup installs it locally with Docling's dependencies, not another copy of the Frisket app. No separate server release is needed, including when running from a source checkout. Later launches start the installed server automatically; models load on first use, which may download model files and take longer. Documents stay on your machine. The server is shared by the app and its background jobs and stops when Frisket exits. Cancelling an action stops waiting and submitting further documents; a conversion already running on the server may finish, but its result is discarded.

This managed installation currently covers Docling only. Existing native engines and operator-configured model servers still work as before. For example, for OCR local/team installs can add Surya 2 with the sidecar's ocr extra and an upstream-supported inference backend. The standard sidecar container and Cloud use dots.mocr instead.

For the models included in the standard sidecar container, build the image yourself by downloading the repo and running the following commands from the repo root.

docker build -t frisket-models:local ./sidecar

Then run the sidecar and connect it to Frisket.

MODELS_TOKEN="$(openssl rand -hex 32)"

docker run -d \
    --name frisket-models \
    --restart unless-stopped \
    -p 127.0.0.1:8500:8500 \
    -e FRISKET_MODELS_TOKEN="$MODELS_TOKEN" \
    -v frisket-models-cache:/models \
    frisket-models:local

export FRISKET_MODELS_URL=http://127.0.0.1:8500
export FRISKET_MODELS_TOKEN="$MODELS_TOKEN"

frisket ./my-workspace

Extras

Clef classification and extraction

The Classify and Extract actions include Clef Flash for an operator-managed model server and Clef (Cloudflare) for Cloudflare Workers AI. Both accept up to 64 output fields: categories with 2–254 labels, booleans, or integer scores from 0 to 10. They can include confidence values; justifications and free-text outputs are not supported. Choose the engine directly; no separate model is needed. One engine answers every field in the request. Clef always selects an answer, even when the input is inconclusive; it does not return “not found.” Confidence is the probability of the selected answer.

Clef Flash also limits the whole request to 512 options (a boolean counts as 2, a score as 11, and a category as its number of labels). Input text and the rendered question schema each have a 65,536-character limit; each question's instructions and each label description have a 4,096-character limit. Dataset context and extraction instructions count toward these limits. Text, questions, and model formatting must together fit its 16,384-token context, so requests within the character limits can still be too large. Shorten the text or instructions, or reduce fields and labels if Flash rejects a request. These worker limits do not apply to hosted Clef.

For Extract, use text source columns and set Citations to Off. Clef does not support citations, image input, free text, arbitrary numbers, lists, or JSON objects. Switching engines keeps existing fields and citation settings visible; incompatible settings must be corrected before running. Choose a generative model for those settings. Extract's instructions, dataset context, required fields and output column names also apply to Clef.

For Clef Flash, install the model server's classify-clef extra and configure FRISKET_MODELS_URL and FRISKET_MODELS_TOKEN as above. It becomes available when the authenticated server advertises clef-flash on /classify in its capabilities. The managed Docling installation does not install Clef Flash.

For hosted Clef, set CLOUDFLARE_ACCOUNT_ID to your 32-character hexadecimal Cloudflare account ID and CLOUDFLARE_API_TOKEN to your Workers AI API token before starting Frisket. The token can also be saved as CLOUDFLARE_API_TOKEN in the project's Settings → Secrets. Hosted classification sends the selected text to Cloudflare and uses Frisket's external-action cost and consent flow.

Local models

Fair warning: some of these aren't yet available without some additional setup.

OpenDocRouter is available in To Markdown, OCR, and OCR Compare. Choose a model and add your API key in the selector, or set OPEN_DOC_ROUTER_API_KEY on the server. It accepts PDF, PNG, and JPEG files. To Markdown sends the original document; OCR and OCR Compare process rendered pages and return plain text with block-level highlights. OCR also supports searchable PDF output when the model returns usable layout data. Text selection alignment is approximate for paragraph-sized blocks. PDF generation requires layout data and uses the existing Latin/Western text-layer composer.

Each provider request accepts up to 50 MB and 500 PDF pages. To Markdown requests exceeding the selected model’s synchronous page limit (at most 50 pages) use OpenDocRouter's encrypted temporary result storage (up to 24 hours); Frisket requests deletion after retrieving and validating the result, or when you cancel. If polling, retrieval, or validation fails, Frisket leaves the job and its results available for reconciliation until the provider's retention window expires, rather than submitting another paid job. Smaller requests do not enable caching. Model names and per-page maximum price estimates refresh through Frisket’s existing daily pricing feed, with a bundled catalog for offline startup. Estimates use the published maximum per page; actual costs use provider-reported usage, including partially failed calls. Requests with unknown costs require confirmation. Requests whose responses are lost are not automatically resubmitted.

It's easy to set up an API key to talk to AI providers like OpenAI, Anthropic, OpenRouter and Gemini. But! You can also do most everything on the privacy of your own machine.

• LM Studio and Ollama are supported out-of-the-box for local LLMs/VLMs
• Local transcription can be powered by Parakeet, Whisper, MOSS, VibeVoice-ASR
• spaCy or GLiNER for local entity extraction
• Plenty of OCR engines like RapidOCR, Indra OCR, Surya 2, dots.mocr, PaddleOCR-VL, Tesseract
• PDF to Markdown conversion with Markitdown, Docling

Indra OCR runs locally on CPU using an approximately 18 MB model. It is by Datalab and uses a restricted model license. It is available for local/self-hosted use, not through Frisket Cloud.

Reclaim unused local files

Stop Frisket before running cleanup. Also stop independently launched workers, local MCP clients, and scripts using the workspace. The command detects a running normal Frisket launcher, but cannot detect every external reader or writer.

frisket cleanup ./my-workspace/my-project.frisket         # inspect only
frisket cleanup ./my-workspace/my-project.frisket --apply # reclaim unused bytes

Cleanup preserves files needed by retained history, evidence, and resumable imports. It does not clear undo history. Byte counts are the sizes of removed files, not filesystem blocks; files still needed by the project remain. This is manual, offline maintenance for local bundle storage, not a hosted/S3 cleanup command. Restart Frisket afterward.

Write a plugin

Plugins can add actions, panels, imports, integrations: pretty much anything. They're easy to make!

frisket plugin init --id example.my-plugin --output ./my-plugin --with action
frisket plugin build ./my-plugin
frisket plugin validate ./my-plugin

Me

Hi, I'm Soma!

I teach data journalism at Columbia's J-School where I run a year-long Data Journalism MS and a nice short summer program. I give a lot of talks on AI and wrote history's friendliest PDF-processing library.

Email me at jonathan.soma@gmail.com.

People say things like "oh buy me a coffee if you like this" but no, I'm more demanding: go look at these poor cats and then donate to my cat rescue.