NativeBIM by NativeCAD

Privacy: what the NativeBIM IFC server sends, and what it keeps

This covers nativebim-mcp, the NativeBIM MCP server for IFC models. It describes what the server software sends and stores. How featureboard.ai stores what it receives is covered by the featureboard.ai terms. Where this statement and the software disagree, the software's behavior is what happens, and this statement is corrected.

In short

• Your models, drawings and files stay on your machine. NativeBIM never uploads them.
• NativeBIM sends four kinds of message, all to featureboard.ai:
  1. daily usage counts;
  2. automatic failure reports;
  3. bug reports that you, or the assistant acting for you, write;
  4. an update check.
• Usage counts, failure reports and the update check do not contain model content, file names, paths, prompts, IFC GlobalIds or element names. A bug report contains the text that you or the assistant type. That text goes through a pattern-based redactor before it is sent. A pattern can miss a value, so a bug report can contain something you did not mean to share (see Bug reports you write).
• License keys are checked on your machine. No network call is made to check a key.
• Every message can be turned off, with two exceptions:
  • on a free license, failure reports are a condition of the license;
  • NATIVEBIM_SANDBOX=1 stops all four from leaving the machine.

The four messages

These three addresses can be changed with NATIVEBIM_TELEMETRY_ENDPOINT, NATIVEBIM_BUGS_URL and NATIVEBIM_UPDATE_URL. Setting the telemetry or bug address to a value that is not a URL, such as local, keeps everything on the machine. Plain http:// is accepted only for addresses on the same machine.

1. Usage counts

There is one message per install, per server process, per UTC day. These are all its fields; the code drops any other field (telemetry/schema.js):

Tool names come from NativeBIM's own list. A name outside that list is counted as _other, and an unknown code as _OTHER. Each field must be a fixed value, one of a closed list of words, a whole number, a date, or match a tight pattern, so the message has no field that could hold free text. A test checks this (tests/mcp/telemetry/schema.test.js).

A day's counts are sent after that day ends, when the server is next running. They stay on disk until then, and are written there every 10 minutes and at shutdown. If an upload fails, the message stays on disk and is retried later, with the wait growing from 1 minute up to 6 hours. If the server refuses a message outright, it is deleted. The queue holds at most 120 messages or 1 MB; past that, the oldest are deleted.

2. Failure reports

These are sent when NativeBIM itself breaks: a tool throws an error it did not expect, or returns an *_INTERNAL code. A refused request is not a failure. That covers a bad argument, a missing license or a selector that matches nothing. Fields:

• fingerprint: a hash that groups reports of the same failure;
• class, severity;
• the tool name, the NB_* code and, where the tool gives one, a short reason code made of lowercase letters, digits, _ and -;
• the error type, for example TypeError;
• title and message: made only from the tool name, the error type and the code;
• stack: up to 12 lines of the program's call stack. Each line keeps only a function name and the name of a NativeBIM .js file. Folder names, and the names of any non-code files, are removed;
• argShape: the names of the arguments and their types, for example {"target":"string","ops":"array[3]"}. Never the values;
• versions, edition, OS and Node version;
• a random install id for failure reports (a UUID);
• count: how many times this failure happened;
• reportedAt;
• on a free license only, the key id and tier: "free".

The error message itself is never sent, because messages can quote values and file names. Before sending, each report is checked field by field. Any field without a rule is dropped. The final bytes are checked once more for anything that looks like a path, file name, URL, e-mail, GlobalId, GUID or IFC record, or that matches a name from the open models. If anything matches, the report is not sent.

Limits:

• one report per failure per 24 hours; later occurrences are added to the next report's count;
• at most 30 failure reports an hour;
• unsent reports wait on disk (1 MB at most) and are retried.

Who sends them (the same rule as NativeCAD):

3. Bug reports you write (bim_report_bug)

These are sent only when bim_report_bug is called. Fields:

• what was typed: title, description, steps, expected, actual, severity, and which bim_* tool is involved;
• context: NativeBIM version, edition, the name and version of the app NativeBIM runs in (as it announced itself when it connected), OS, Node version, kernel version;
• context.modelCounts: counts only (models, entities, elements, storeys, spaces, types);
• guard: how many values the redactor replaced, by kind (counts only);
• contact: your e-mail address, only if includeEmail: true is passed. It is read from NATIVEBIM_REPORTER_EMAIL;
• reportedAt;
• on a free license only, the key id and tier: "free".

Before anything is sent, these are replaced in the typed text with [redacted:<kind>]:

• any name NativeBIM knows from the open models: file, project, story, space and element names, and the folder names in the paths of the open files;
• the account name, home-folder name, host name (whole, and the part before its dot) and working-folder name of this machine, and each word of four letters or more in an account or home-folder name made of several words (Jane Smith gives Jane and Smith). Names that are ordinary words, such as Users, Documents or admin, are left in place;
• file paths and file names, including folder and file names that contain spaces;
• URLs and e-mail addresses;
• IFC GlobalIds, GUIDs and Revit UniqueIds;
• raw IFC records such as #45=IFCWALL(...), including a record that wraps over several lines;
• license keys.

Names are replaced before the pattern rules run, longer names before shorter ones, so a known path or file name goes as a whole. These are pattern-based rules. They can miss a value, for example a name that is not in the open models and does not look like a path, file name, identifier or key, or the leading words of a file name with spaces that is typed in a sentence rather than as part of a path. The rest of the text is sent as typed. Do not paste model content or prompts into a bug report.

A report may carry a small repro, such as a few lines of IFC, only with your consent. When your AI app supports MCP elicitation, NativeBIM asks you directly (showing what would be attached) and your answer decides; otherwise the assistant must pass reproConsent: true with it. That is the one way model content can reach us, and it is your choice. Paths, file names, URLs, e-mail addresses, license keys and this machine's names are still replaced in it; IFC records, GlobalIds and model names are kept, because they are the repro.

Limits:

• the same report twice within 10 minutes is filed once;
• at most 6 reports an hour; more are queued and sent later;
• a report that cannot be sent is queued and retried.

A copy of every report, after redaction, is kept in ~/.nativebim/bugs/bugreports.jsonl. The receipt shows that location with your home folder written as ~ (%USERPROFILE% on Windows), so the assistant does not see your account name; the file itself sits at the full path on your disk. The receipt gives the report id. Quote it to support@featureboard.ai.

4. Update check

This is a request for the public file latest.json, at most once every 6 hours while the server is in use. The request has no body and no identifiers. Other requests happen only when something changed:

• the new release, when an update installs;
• the signed revocation list, when its contents change.

Whether this install is in a staged rollout is decided on the machine, from a random id in ~/.nativebim/update/state.json. That id is never sent. A release installs only if:

• its manifest is signed with NativeBIM's release key;
• the download matches the published checksum;
• no tool call is running.

NativeBIM keeps the previous version and goes back to it if the new one fails to start three times.

License checks

Keys (CSLV1-...) are checked on the machine. No network call is made to check one.

NativeBIM looks for a key in three places:

1. NATIVEBIM_LICENSE_KEY;
2. ~/.nativecad/license.json, NativeCAD's file, which NativeBIM reads but never changes;
3. ~/.nativebim/license.json, written by bim_license activate and readable only by you (mode 0600).

The key itself is never sent anywhere and never written to a log. The key id is the last 16 characters of the key. It is shown in bim_license status and is sent only where this page says so. featureboard.ai issued the key, so it can match a key id to the license it issued.

The list of revoked key ids is signed. NativeBIM reads it from disk. If the list is missing or out of date, no key is blocked because of that.

What is never sent

• Model content: IFC entities, geometry, coordinates, property values, drawings, schedules. The only exception is a repro you attach with consent.
• File names and file paths.
• Prompts, and anything else you or the assistant typed into a tool call. Bug reports are the exception: what you type into bim_report_bug is sent, after the pattern-based redactions described above, which can miss a value.
• IFC GlobalIds, GUIDs, element, space, story and project names.
• Tool arguments (only their names and types, in failure reports), tool results, and error messages.
• Your license key, your name or e-mail. The one exception is the e-mail, sent only when you pass includeEmail: true with a bug report.

Identifiers

What stays on your machine

Everything is under ~/.nativebim/, or under the folder named by NATIVEBIM_STATE_DIR (or its alias NATIVEBIM_HOME) if either is set; NATIVEBIM_STATE_DIR wins when both are. You can delete any of it.

The audit log is a local record of every committed edit, undo, redo and IFC save, so a professional can show what was changed and when. Each line holds the time, the tool name, a hash of the edit instructions (not the instructions), a label for who made the change (NATIVEBIM_AUDIT_ACTOR, or local; your account name is not read), the server process id, the hash of the model file before and after, counts of records added, changed and deleted, the validation result, tool versions, and the message you gave the edit with paths, file names, e-mail addresses and IFC ids removed. It holds no file contents, element names, file names or paths. Lines are chained by hash so an edited or removed line can be detected. Nothing in it is ever sent: the update check, usage counts and bug reports do not read it. The design-log export writes it to a file you choose. Keep it as long as you need to; NativeBIM never deletes it, and turning telemetry off does not touch it.

Turning usage counts off with the telemetry setting also deletes the waiting messages, the counts and the install id.

Turning things off

When NativeBIM is started by its own test suite (node --test), nothing is reported unless NATIVEBIM_TELEMETRY=on is set, and then only to an address on the same machine.

A server started with usage counts turned off writes no telemetry files at all.