Pseudonymizer plugin

The built-in Pseudonymizer plugin keeps personal data away from the provider of a model. In a pseudonymized project, what its sessions send (prompts, files read, tool results) goes out with pseudonyms in place of names, places, phone numbers and identifiers: [PERSON_001], [LOCATION_002], [IBAN_001]. The real values stay on your computer, and the answers, the tool calls and the files the agent writes get them back.

Detection runs on your computer, in CodeAlta: patterns and checksums for identifiers, and for names a local model of the language of your documents. Nothing is sent to a service to find what to hide.

The plugin is for CodeAlta Desktop. In CodeAlta TUI, the sessions of a pseudonymized project send nothing.

How it works

sequenceDiagram
    participant You
    participant CodeAlta as CodeAlta (your computer)
    participant Provider as Provider of the model
    You->>CodeAlta: "Summarize the letter of Sophie Martin"
    CodeAlta->>Provider: "Summarize the letter of [PERSON_001]"
    Provider->>CodeAlta: "[PERSON_001] asks for…"
    CodeAlta->>You: "Sophie Martin asks for…"
  • One value, one pseudonym. A value keeps its pseudonym for the whole session, so the model can follow who is who. Pseudonyms are numbered by kind in the order they appear.
  • The model is told. It is asked to copy each pseudonym exactly, never to guess what it stands for.
  • Your tools get the real values. A command, a file edit or a search on your computer runs with the real values. The tools that send their arguments away from your computer (web search and fetch, a browser page) get the pseudonyms: see Tools that get pseudonyms below.
  • What you see is real. The conversation, its history and the files keep the real values. Only the requests to the provider hold pseudonyms.
  • The table stays here. The pseudonyms of a session and their values are stored encrypted with your account's key on this computer (DPAPI, the Keychain or libsecret).

Pseudonymize a project

  1. In the Explorer, open the menu of the project and choose Pseudonymize the project….
  2. Choose the Language of the documents (see below).
  3. Optionally, open Names and words:
    • Names always hidden: the parties of the file, hidden wherever they appear, even where nothing finds them. One per line. Use a prefix such as LOCATION: Montauban to choose the kind of pseudonym; older prefixes such as LIEU: still work.
    • Words never hidden: what was taken for personal data by mistake. One per line. Names always hidden takes precedence in new text.
  4. Choose Pseudonymize. The project shows a padlock, and its next requests are pseudonymized.

The same menu item, Pseudonymization…, changes these choices later or stops pseudonymizing the project. The choices are saved in the project's .alta/config.toml.

Languages and models

Each project chooses the language of its documents, which chooses the model that finds names. A model is downloaded once, the first time a project needs it, and runs on your computer.

Language What it finds
French Names of people, places and organizations in French, and the identifiers of France (NIR, SIRET…).
English Names of people, places and organizations in English, and the identifiers of the UK and the US (NI, NHS, SSN…).
Several languages Ten languages, French and English among them, and the identifiers of France, the UK and the US.
Identifiers only Emails, phone numbers, IBAN, social security and company numbers of France, the UK and the US. Names are not looked for, and nothing is downloaded.

Settings > Plugins > Pseudonymizer lists the languages with their model and its size, and deletes a model or downloads it again. A Custom model setting takes the folder of a token classification model of yours (ONNX, with its tokenizer and a ner-model.json that describes it). Descriptions that use older French kinds, such as PERSONNE, still work without edits or a new download.

What is hidden

People, contact details (emails, phone numbers, postal codes) and identifiers (bank accounts and cards, social security, passport and tax numbers…) are always hidden. Settings > Plugins > Pseudonymizer decides the rest, for every pseudonymized project:

Setting By default
Places Hidden: streets, towns, regions and countries the model finds.
Organizations Hidden: companies, firms and administrations the model finds.
Company numbers Hidden: SIREN and SIRET, Companies House numbers, EIN.
Every date Off: only dates of birth are hidden, the dates of acts and meetings are sent.
Web addresses Off: links to documentation, a repository or a site are sent.
Words never hidden CodeAlta, CodeAlta Desktop, noreply@codealta.io, and the technologies CodeAlta's own instructions name (CUDA, ROCm, Vulkan, Metal, EPUB, MarkItDotNet, Unicode, CI, PowerPoint, io): add the names of products or of your firm that name no one. A project whose party has one of these names lists it in Names always hidden.
Tools that get pseudonyms The web tools and the pages of a browser. * stands for any characters.

See what was hidden

  • In the composer, the Pseudonymizer item tells how many values the session hid. Choose it to list the pseudonyms the provider read, their kind and their value, and how many times each was sent.
  • Pseudonymizer: try a text, in the command search, shows what a text would become, with the pseudonyms of your project's language, without sending anything.

Images, PDFs and documents

  • Text read by a tool (a file, the output of a command, a document converted to text) is pseudonymized like a prompt.
  • A PDF attached to a prompt is sent as its text, pseudonymized. A PDF whose text cannot be read is not sent.
  • An image is not sent: the model gets a note that an image was left out, since the text in a picture cannot be checked.

When nothing is sent

A pseudonymized session sends nothing it could not pseudonymize. The turn stops with a message that says why, when:

  • the Pseudonymizer plugin is off;
  • the model of the project's language is not downloaded yet: the session's Pseudonymizer item offers the download;
  • the model cannot run, or a part of the request cannot be analyzed;
  • the project's .alta/config.toml cannot be read, so CodeAlta cannot tell whether the project is pseudonymized;
  • the provider runs an agent of its own that reads files itself (Claude Code, Codex CLI, an ACP agent): CodeAlta does not see what that agent sends. Choose a provider whose requests CodeAlta sends itself, such as a provider with an API key.

Limits

Detection finds most personal data, not all of it. A name the model misses, a rare spelling or the context ("the mayor of a small town") can still tell who someone is. Add the names that matter to Names always hidden, and check with try a text on a sample of your documents.

Pseudonymization covers the requests CodeAlta sends for the sessions of the project on this computer. It does not cover:

  • A session run in the cloud of a provider. Its task and the messages you send to it go to that cloud as they are.
  • Another machine. A session moved to another machine, or a sub-agent started on one, is not pseudonymized there: that machine does not know the project is.
  • A session of another project. What a pseudonymized session sends to it goes to its provider with the real values, and it reads the real values when it searches or reads the history of a pseudonymized session.
  • The other tools. A tool that is not in Tools that get pseudonyms gets the real values, also when it sends them away from your computer: a browser tool that fills a form or runs a script in a page, a tool of an MCP server. Add the ones you use to the setting.
  • A provider that runs an agent of its own. Its turns are stopped, but not what it does without a turn: its own compaction and Remote Control.
  • What is not text of the conversation. The descriptions of the tools and the names of the fields of their arguments are sent as they are.
  • A text already sent. It is sent again as it went the first time: a name you add to Names always hidden later is hidden in new texts only.
  • The lists of the project. Names always hidden and Words never hidden are saved in clear in the project's .alta/config.toml, beside its documents.