innernet-cli

Build your own Innernet: contacts in, a queryable graph out. Runs on your machine, against your database, from your own exports.

npx create-innernet my-innernet
cd my-innernet && pnpm install

pnpm innernet doctor                          # check db, apply schema
pnpm innernet ingest ~/Downloads/contacts.vcf # vCard or CSV
pnpm innernet resolve                         # merge duplicates
pnpm innernet stats                           # what you've got
pnpm innernet mcp --install                   # hand it to your agent

Commands

| Command | What it does | |---|---| | doctor | Check the database, apply the schema, report which optional keys are missing | | ingest <file> | Import a .vcf or .csv. --dry-run previews without writing | | resolve | Merge duplicates. Proposes the ambiguous ones; --apply merges them too | | stats | Shape of your graph | | search <query> | Find someone by name, company, title, or handle | | forget <name> | Delete someone and tombstone them | | mcp | Print the MCP config, or --install to write .mcp.json |

enrich, index, serve, ask and connect are not in v0.1. They print what is missing and exit 2, so a chained script stops rather than believing the step succeeded.

Setup

One environment variable is required:

INNERNET_DB_URL=postgres://user:pass@localhost:5432/innernet

Put it in .env next to where you run the command, or export it. Anything that speaks Postgres works — local Docker, Supabase, Neon, RDS. doctor creates the schema on first run.

How resolution works

Two passes, and the split is the point.

Pass one merges on a shared email or phone — anchors strong enough to act on unattended. Pass two finds same-name and same-name-plus-company pairs and proposes them without touching anything, because automatic merging is confident about the easy cases and wrong about the interesting ones. Collapsing two real humans into one costs far more than leaving a duplicate in place.

Ingest is idempotent and runs in a single transaction, so re-importing the same file updates rather than duplicates, and a failure partway through writes nothing at all.

Removal is real

pnpm innernet forget "Sarah Chen"

Deletes the record and writes a tombstone, so a later re-ingest of the same contacts file cannot quietly bring them back. Honor the request the first time.

House rules

This builds a database of real people who did not opt in.

• Keep it local. Do not publish it, sell it, or put it on a public URL without auth.
• Profile professional lives, not private ones.
• Never use it for a decision it is not lawful for — hiring, housing, credit and insurance are regulated in most jurisdictions and this is not a compliant consumer report.
• Do not infer sensitive traits.

Full guide: https://jointheinner.net/build-your-own

MIT