Prodigy
Prodigy is a downloadable annotation tool for NLP, LLM and computer vision work, built by the makers of spaCy. It runs entirely on your own machines and is sold as a one-time perpetual license rather than a subscription.
What is Prodigy?
Prodigy is a downloadable annotation tool published by ExplosionAI GmbH in Berlin, the team behind the open-source NLP library spaCy. It ships as a Python package installed with pip using a license key, and it bundles the web application in which the annotation work is done. The publisher presents it as "an extensible annotation tool that gives you a new way to build custom AI systems", the idea being to define a classification scheme with real-world examples rather than just prompts and let models assist along the way, with no machine learning experience required.
The range of tasks is broad: named entity recognition, span categorization, text classification, part-of-speech tagging, dependency parsing, coreference resolution, computer vision, audio and video annotation, transcription, speaker diarization, and model training and evaluation. Six feature pages organize this into Information Extraction, Language Model Training, Computer Vision, Audio and Video, Prompt Engineering and Custom Workflows, while six industry pages address Banking and Finance, Healthcare and Biomedical, Media and Content Creation, Legal and Insurance, Conversation and Insights, and Research and Education.
The extension model rests on recipes: Python functions decorated with @prodigy.recipe that describe a complete workflow, from loading the data to putting a model in the loop and saving the results. Active learning is built in for named entity recognition, text classification, tagging and parsing, so the model proposes and the annotator corrects.
Privacy is the central argument. Prodigy runs offline, including on a machine isolated from the network, and never contacts the publisher or any third party. Storage is the customer's choice, with SQLite by default, MySQL or PostgreSQL, plus JSON export, a Python API and a command-line interface. The publisher claims more than 10,000 developers and researchers as users and announced more than 500 corporate customers in 2021; the logos displayed on the home page, among them Bayer, KPMG, Associated Press, Munich Re, Microsoft, The Guardian, Barings and BASF, are presented as users rather than partners or integrations.
The stated limits are equally explicit. Prerequisites are Python 3.8 or later on macOS, Linux or Windows, and spaCy v3.1 or later for the recipes that depend on it. Note that the same publisher also produces spaCy and Ellf, a separate beta product promoted by a banner across the site.
What it does
- Annotate text, images, audio, video and relations in a web app running on your own machine
- Correct the predictions of a large language model instead of labeling from scratch
- Train and evaluate models from the annotations collected, without leaving the command line
- Develop and compare prompts through blind A/B tests and prompt tournaments
- Pick from 23 annotation interfaces, including multiple choice, HTML blocks and conflict resolution
- Route tasks between annotators, review disagreements and measure annotation quality
- Write your own Python recipes and your own HTML, CSS and JavaScript interfaces
When to use Prodigy / When not to
A quick filter to help you decide if Prodigy is the right fit.
When to use Prodigy
- Data science and NLP teams that train their own models instead of calling a hosted API
- Organizations working with sensitive medical, financial or legal material, including air-gapped environments
- Academic researchers at degree-granting institutions, who can request a free interim research license
- Freelance developers and AI consultants, whose Personal license covers unlimited commercial use
- Python-literate teams that want to script annotation itself, with custom recipes and purpose-built interfaces
When not to use Prodigy
- Non-technical teams: the documentation assumes basic familiarity with the Python programming language and the command line
- Buyers looking for a hosted SaaS annotation platform, since there is no version operated by the publisher
- Individuals who want to test before buying: the trial is a hosted VM granted to companies and organizations only, and there is no permanent free plan
- Public or private entities primarily engaged in military, law enforcement, intelligence or national security work, to whom the publisher refuses to supply the software (regulatory agencies and tax authorities are explicitly exempted)
- Teams that would host the app on a public server, redistribute or sublicense it, or share one seat across several machines at the same time, all of which the license forbids
How to use Prodigy
A typical end-to-end flow, from setup to results.
- Buy a license from the pricing page, operated through the SendOwl store, and receive a license key and an order ID prefixed with #EX
- Install the package with pip from the publisher's PyPi server, using the license key; this has been the delivery method since v1.11.0 in August 2021, and older download links are inactive
- Install a spaCy pipeline separately if your recipe needs one, for example python -m spacy download en_core_web_sm
- Start a recipe from the command line, for example prodigy ner.llm.correct news_articles ./config.cfg ./news.jsonl
- Annotate in the browser, where the web app serves the examples one at a time
- Train a model from the annotations collected, for example prodigy train ./information_extraction --ner news_ner --textcat news_textcat
- Choose a deployment mode: local machine, optionally shared through ngrok, a virtual machine rented from a cloud provider, or a Docker container
- Protect the instance with HTTP basic authentication through the PRODIGY_BASIC_AUTH_USER and PRODIGY_BASIC_AUTH_PASS environment variables, or with OpenID Connect SSO on a Company license
- Schedule your own database backups, for example sqlite3 ~/.prodigy/prodigy.db '.backup /path/to/backup' run from CRON
- Extend the tool when needed: write a Python recipe decorated with @prodigy.recipe, combine interface blocks, or plug your own front end into the documented REST API
Pros & Cons
Pros
- Maximum confidentiality: nothing leaves the customer's machines, and the tool can run in a network-isolated environment
- One-time purchase with a lifetime license, so no subscription and perpetual access to the versions covered by the order
- No lock-in: the models produced, the data formats and the storage back ends are open, with JSON export
- Real extensibility: Python recipes, purpose-built interfaces, and a REST API whose source is shipped with the product
- Native integration with spaCy, from the same publisher, as well as Hugging Face, OpenAI, Cohere and Whisper
- Active learning reduces the number of human decisions needed to reach a usable model
- Free research license for researchers at degree-granting institutions, and flexible purchasing by purchase order, invoice or resellers such as SHI, SoftwareOne and QBS Software
Cons
- Setup requires Python and the command line, which puts it out of reach of non-technical teams
- No permanent free plan, and the trial is a hosted VM reserved for companies and organizations, not offered to individuals
- High entry cost for a team: 490 USD per seat sold in packs of 5 seats, a minimum ticket of 2,450 USD
- Updates are included for 12 months only; beyond that, a further year has to be bought separately
- One seat allows only one machine at a time, and the whole infrastructure (virtual machine, database, backups, authentication) is the customer's responsibility
- Minimal legal documentation: a summary privacy policy, no mention of the GDPR, no data processing agreement and no list of subprocessors; the terms are dated January 29, 2021 and are governed by English law although the publisher is German
- No mobile app, no interface language declared other than English, and dead internal links observed on September 2, 2026, with /teams/recipes/train returning a 404 while /features and /teams serve the home page
Pricing & Plans
There is no free plan. Prodigy is sold as a one-time purchase rather than a subscription: the lifetime license is paid for once and used indefinitely. The lowest price point is the Personal license at 390 USD per lifetime license, excluding tax. The Company license is 490 USD per seat, excluding tax, sold in packs of 5 seats, which sets a minimum order of 2,450 USD. Both include 12 months of free updates, after which an additional year of updates is purchased separately. VAT is added to orders placed from the European Union and the United Kingdom, with reverse charge for VAT-registered businesses outside Germany and no tax outside the EU and the UK. Payment is accepted by card, validated by Stripe, as well as by PayPal or bank transfer, with purchase orders and invoices available for Company licenses; the resellers SHI, SoftwareOne and QBS Software also distribute the product. A free interim research license is granted on request to researchers at degree-granting institutions. The structured price fields of this entry (starting price, currency code and billing unit) are deliberately left empty because the available billing units describe recurring charges only and none of them fits a perpetual one-time purchase; the exact amounts are stated here instead.
- Personal — 390 USD per lifetime license
- excluding tax. Aimed at freelancers
- independent developers and hobbyists. Includes 12 months of updates
- the installer
- the web app and the plugins
- HTTP basic authentication
- community forum support
- and unlimited personal and commercial use. The license is issued to an individual and is not transferable.
- Company — 490 USD per seat
- excluding tax
- sold in packs of 5 seats
- a minimum order of 2
- 450 USD. Aimed at startups
- data science teams and enterprises. Includes 12 months of updates
- the installer
- the web app and the plugins
- Research license — free
- interim
- granted on request by email. Reserved for researchers at degree-granting institutions.
Data, GDPR & hosting
A consolidated view of how Prodigy handles your data.
GDPR overview
There is no mention of the GDPR anywhere on the site: a search across every collected page returns zero occurrences of GDPR, RGPD or DSGVO. The publisher is nevertheless established in the European Union, as ExplosionAI GmbH in Berlin, so no Article 27 representative is required or named. No data protection officer is appointed and no dedicated channel for data subject rights is published; the privacy policy simply ends with the generic address contact@explosion.ai. That policy is summary in nature, stating no legal basis, no retention period and no list of rights, beyond a clause declaring that the services do not address anyone under the age of 13. No data processing agreement is published or offered. The de facto position is structural: because the software is self-hosted, the publisher never processes the data being annotated. The terms themselves are governed by the law of England and Wales.
Who owns the data?
Prodigy is self-hosted, so annotations and trained models stay on the customer's own infrastructure. The home page states that "The models you produce are yours as well, with absolutely no lock-in", and the privacy policy explicitly places the software outside its own scope: Prodigy "is entirely self-hosted and does not connect to ExplosionAI GmbH's servers or any other third-party providers". Data lives in a database the customer chooses (SQLite, MySQL or PostgreSQL) and exports to JSON. The privacy policy covers only the prodi.gy website, the online store and the support forum; content uploaded there grants the publisher a non-exclusive, royalty-free, worldwide license, limited to the purpose of providing the service.
Reuse rights
The downloaded software transmits nothing. No telemetry is declared, Prodigy can run on a machine with no network connection, and the customer therefore reuses annotations and models freely, for any purpose, without asking the publisher's permission. Collection is confined to the website, the store and the support forum, where the publisher records Log Data (IP address, browser version, pages visited, date and duration of the visit) and cookies, together with personal details such as name, telephone number and postal address supplied by the user, in order to contact or identify them. Third-party service providers may access that website data to perform tasks on the publisher's behalf, and no list of them is published; card payments are validated by Stripe and the store is operated through SendOwl. One qualification matters: recipes that integrate with an external vendor such as OpenAI send the data being annotated to that vendor's servers. The documentation says so plainly and advises opting out of those recipes when the data is sensitive.
Data retention & training
Hosting summary
Nothing is hosted by the publisher. Prodigy runs on infrastructure the customer provides and controls, in one of three documented modes: a local machine, optionally shared over ngrok, a virtual machine rented from any cloud provider, or a Docker container; a fully on-premises deployment is equally possible. The database is the customer's choice, with SQLite by default in the file ~/.prodigy/prodigy.db, or MySQL or PostgreSQL, and backups are the customer's responsibility, the documentation suggesting a scheduled sqlite3 backup run from CRON. As a result ExplosionAI GmbH declares no hosting country and no hosting region, an absence that is consistent since it hosts nothing: the jurisdiction covering annotation data is whichever one the customer chooses. The only exception is the publisher's own perimeter, meaning the marketing site, the online store and the support forum, whose hosting is not documented; the IP address resolved for the domain is an AWS Global Accelerator anycast node, which identifies a network edge rather than the real origin.
Things to keep in mind
Risks and trade-offs to weigh before adopting Prodigy.
- Domain signals are misleading: prodi.gy was registered on June 30, 2017 while the first Wayback capture dates from March 2, 2012 and the site fingerprints differ until 2016, so the archive measures a previous owner; the .gy extension is the Guyana ccTLD used as a pun and indicates no presence there, the publisher being German
- Three products from the same publisher are easy to confuse: spaCy, which is open source, Prodigy, described here, and Ellf, a beta assistant promoted by a banner at the top of every page. Prodigy Teams, a hosted version announced in a 2021 blog post, does not exist as a product on the site, and the home page FAQ link to /teams/recipes/train returns a 404
- The logos shown on the home page, including Bayer, KPMG, Associated Press, Munich Re, Microsoft, The Guardian, Barings and BASF, are claimed users, not integrations and not partners
- The REST API is the API of the software the customer self-hosts, not an endpoint operated by the publisher: there is no hosted service to call
- Recipes that call an external vendor such as OpenAI or Cohere send the data to that vendor's servers, which suspends the offline guarantee for those specific workflows; the documentation states this explicitly
- The terms are dated January 29, 2021 and are governed by the law of England and Wales although the publisher is a German GmbH, so the contract text has not been revised in more than five years
- Beneficial ownership is inferred rather than registered: the German commercial register does not publish beneficial owners and the Transparenzregister is not freely searchable, so the named individuals come from the cap table the publisher published itself
Setup & Integrations
Technical difficulty
Moderate for whoever installs it, negligible for whoever annotates. Installation is a pip command with a license key, like any Python library, and requires Python 3.8 or later on macOS, Linux or Windows, plus spaCy v3.1 or later and a separately downloaded pipeline for the recipes that need one. The publisher states the prerequisite openly: basic familiarity with Python and the command line. Once a task is configured, annotating in the browser requires no programming. Beyond a single workstation, the customer provisions the virtual machine or container, the database, the backups and the authentication.
Deployment
Integrations
Behind Prodigy
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What makes Prodigy different from other annotation tools?
Is my data really private?
What models can I train with it?
How technical do I need to be?
Is there a free trial?
How much does Prodigy cost?
What happens when the 12 months of updates expire?
Which cloud providers are supported?
Is there an offer for universities?
Are there buyers the publisher refuses to sell to?
Should you pick Prodigy?
Prodigy is a deliberately narrow tool. It addresses developers and data scientists who build their own models, and it assumes they are at ease with Python and the command line. The central trade-off is stated openly by the publisher: full autonomy and confidentiality in exchange for the infrastructure and the technical skill the customer has to supply. Nothing is hosted by ExplosionAI GmbH, so annotation data and trained models never leave the customer's machines, which carries real weight for medical, financial, legal or air-gapped work and also explains why no hosting country or region is declared.
The commercial model runs against the current: a one-time perpetual purchase at 390 USD for a Personal license, or 490 USD per seat in packs of 5 for a Company license, rather than a subscription, from a publisher that describes itself as profitable and independent since 2016 and no longer raising venture capital. Updates, however, are included for 12 months only, and there is no permanent free plan; the trial is a hosted VM granted to companies and organizations on request, never to individuals.
Two reservations deserve attention before buying. The legal documentation is minimal: a summary privacy policy, no mention of the GDPR, no data processing agreement and no list of subprocessors. That weighs little for self-hosted software the publisher never touches, but it remains worth checking for the store and the support forum, which are the only things the policy actually covers. And one usage restriction is unusual enough to be checked first: the publisher refuses to supply the software to entities primarily engaged in military, law enforcement, intelligence and national security work, with regulatory agencies and tax authorities explicitly exempted.
- Choosing a selection results in a full page refresh.
- Opens in a new window.