What is iKnow?
Since macOS 13, two background daemons, knowledgeconstructiond and intelligenceplatformd, have built a knowledge graph of the Mac's owner: who you know, how the system decided they are related to you, who you write to, whose faces recur in your photos. According to the Eighth Rank's write-up, it lives in ~/Library/IntelligencePlatform as 51 plaintext SQLite databases, the main one a triple store with 181 entity classes and 278 predicates, built from Contacts, Calendar, Photos, Messages, Mail and other sources. Apple documents almost none of it.
iKnow opens that store and shows it in your browser: what was recorded, what was inferred and with what confidence, and which other parts of the system read it. It is a single Python standard-library server with no dependencies. It reads the graph and the four sources it resolves against (Messages, Photos, Calendar and Contacts) in place, with every database opened read-only. A MANUAL_FOR_YOUR_AGENT.md file explains the same tables to Claude Code or Codex for plain-language questions.
Is iKnow worth it?
iKnow's value is that it shows something nobody else does. Apple ships a one-line man page for the daemon that builds the graph and no user-facing account of what it infers, so a viewer that lays out who your Mac thinks your family is, and why, is useful to privacy-minded users and to journalists alike. The researchers at the Eighth Rank published their write-up on 28 October 2026; the repository had 605 stars, 28 forks and four commits by 13 November, the latest on 8 November.
The security design is sensible for what it touches. The server binds 127.0.0.1 with no flag to change it, every page and endpoint needs a random token printed at start-up, a Content-Security-Policy blocks off-machine requests, and every database handle is opened read-only. The README says nothing is written or logged. The token keeps out other accounts and web pages on the same Mac, but the README notes it is not a sandbox against a program already running as you.
The permission it needs is the heavy part. The terminal you launch it from must have Full Disk Access, which covers everything else you run from that terminal too, not just iKnow. The README tells you to turn it off after each session, and the server reminds you when it stops. The agent manual is the riskier path: unlike the server, an agent sends what it reads to its model provider, and the manual tells it to ask you first.
The catch: it is tested only on macOS 26.6.2, 26.7.1 and 27.0.1, with some pages incomplete on 27, and Apple changes these stores between releases. The README itself says the tool is fully LLM-generated and may not be 100% accurate, so treat its interpretation of a column as a reading, not Apple's documentation.
Our take: run it once, look at what your Mac has inferred about the people around you, then revoke the permission; it is an inspection tool, not something to leave running.
Where it earns its place
A privacy-conscious Mac user
See what my Mac has decided about my family and close contacts
Clone the repository, grant Full Disk Access to Terminal, run `python3 server.py` and open the printed link. The pages list people, inferred relations and their confidence. It costs nothing; switch the permission off afterwards.
A technology journalist investigating on-device AI
Document what Apple Intelligence infers locally, with evidence
Run iKnow on a test Mac, or point `--bundle` at an extracted copy of the stores, and cite the tables it reads. Its interpretations are LLM-written, so check key columns against the SQLite files directly.
A developer who uses Claude Code on their own Mac
Ask plain questions about the graph without learning its schema
Give the agent MANUAL_FOR_YOUR_AGENT.md and ask who the Mac thinks your family is. The agent needs the same Full Disk Access and sends what it reads to its provider, which the manual makes it confirm first.
Before you commit
- Setup
- `git clone`, then `python3 server.py`; the terminal app needs Full Disk Access and a full restart. Start-up prints ok or DENY beside each store.
- Integrations
- Reads ~/Library/IntelligencePlatform plus Messages, Photos, Calendar and Contacts; a manual for Claude Code, Codex or any agent with a shell.
- Data and privacy
- Everything stays on the Mac: loopback only, read-only database handles, a fresh token each start, and nothing written or logged according to the README.
- Limits
- Tested on macOS 26.6.2, 26.7.1 and 27.0.1 only; some pages are incomplete on macOS 27; interpretations are LLM-generated and may be wrong.
What’s good
- Shows inferred relationships and confidence scores Apple does not surface
- Loopback only, token-gated and strictly read-only
- No dependencies beyond Python 3
- Agent manual asks before sending graph data to a model provider
What’s not
- Needs Full Disk Access for the terminal you launch it from
- Tested on three macOS releases; pages can break on others
- Self-described as fully LLM-generated and possibly inaccurate
- Tiny project with four commits
Best for
Mac users, privacy researchers and journalists who want to see what macOS has inferred about them and the people in their contacts.
iKnow alternatives
iKnow: common questions
- Is iKnow worth it?
- iKnow is a small, careful window onto data your Mac never tells you it keeps, worth twenty minutes for anyone curious about Apple's on-device profiling, provided you switch Full Disk Access off again afterwards. We rate it 3.6 out of 5.
- How much does iKnow cost?
- Free and open source (Apache-2.0).
- What are the downsides of iKnow?
- Needs Full Disk Access for the terminal you launch it from; tested on three macOS releases; pages can break on others; self-described as fully LLM-generated and possibly inaccurate; tiny project with four commits.
- Who is iKnow best for?
- Mac users, privacy researchers and journalists who want to see what macOS has inferred about them and the people in their contacts.
- Is iKnow safe for company data?
- Everything stays on the Mac: loopback only, read-only database handles, a fresh token each start, and nothing written or logged according to the README.
- What are the best iKnow alternatives?
- The closest alternatives we have reviewed are screenpipe.
Answers checked Nov 13, 2026.