people-context

English | 简体中文

Your agent already remembers your codebase. Now it can remember your people.

people-context is a local-first MCP server and CLI that gives AI agents durable memory about the people in your life: who someone is, how you know them, what you last agreed, and how they like to be talked to. One SQLite file on your machine. No account, no cloud, no network calls.

OpenSSF Scorecard

pctx demo: seed a fictional dataset, list people, and print a brief

Why

Ask an assistant "how should I approach Priya about the reporting delay?" and it has nothing: it does not know which Priya, that she is your counterpart at a partner org, that you agreed a new deadline last week, or that she prefers a short email over a call. That knowledge lives in your head, your inbox, and a notes file the agent cannot see.

people-context keeps it in one place the agent can query through narrow tools:

• Who is this? Explainable name resolution over names, nicknames, aliases, and handles. Two Priyas come back as two candidates with a match reason, never a silent guess.
• What do I know? Relationships, organisations and roles, durable facts, concise interaction summaries, traits, reminders, and a per-person timeline, each disclosed only as far as the request needs.
• How do I talk to them? Communication guidance grounded in recorded traits, past friction, open follow-ups, and your own written philosophy.
• Who has gone quiet? Stale-relationship and upcoming-date reports over what is already stored.
• Get data in safely. Email, mbox, vCard, calendar, LinkedIn, Outlook, and WhatsApp exports are staged as reviewable candidates. You approve what gets recorded; raw source content is never kept.

It is opinionated about trust: writes are audited, forget is a real delete, sensitive records sit behind an operator-only gate that a prompt cannot open, and ordinary commands never touch the network.

Demo

A packaged fictional dataset is the fastest way to see identity resolution, graph traversal, and bounded context without touching real data:

uvx --from people-context pctx demo --reset

The demo always writes its own dedicated database at {XDG_DATA_HOME or ~/.local/share}/people-context/demo.db. It ignores --db, PEOPLE_CONTEXT_DB, the config file, and workspace discovery, and --reset replaces only that file plus its -wal/-shm companions, so a real database is never read or modified. Seeding writes audited fictional people, handles, affiliations, facts, interactions, and a connected relationship graph, then prints the path-targeted server command and concrete tool calls that use the ids it just created:

Demo database: /home/you/.local/share/people-context/demo.db
Start MCP server: people-context-mcp --db /home/you/.local/share/people-context/demo.db
resolve_person {"query": "Amina Hassan"}
get_relationship_graph {"person_id": "<amina-id>", "depth": 2}
find_connection {"person_a": "<self-id>", "person_b": "<sofia-id>"}

Person ids are generated per seed, so the printed values differ from the placeholders above. Start the printed server command in an MCP client and run the printed calls verbatim. See docs/cli.md.

Quick start

Upgrading from a release before the shared ~/.pctx/people.db default? Before any client starts the new version, inventory and pin each client's existing database as described in Upgrading to the shared default.

Requires Python 3.11+ and uv. Pick your client; each is one step.

claude plugin marketplace add JinyangWang27/people-context
claude plugin install people-context@people-context-plugins

codex plugin marketplace add JinyangWang27/people-context
codex plugin add people-context@people-context-plugins

{
  "mcpServers": {
    "people-context": {
      "command": "uvx",
      "args": ["--from", "people-context", "people-context"]
    }
  }
}

openclaw plugins install clawhub:openclaw-plugin-people-context

uv tool install people-context
pctx init        # seed your own record, optionally import a vCard, then connect a client
pctx --help

Then try, in your agent:

Who is Amina?

Remember that Amina from Open City Lab prefers short emails and hates surprise calls.

What should I know before my meeting with Daniel tomorrow?

The second one is a single remember tool call: the name is resolved, the person is created only if nobody matches, and the affiliation and preference are recorded in one audited transaction. Ambiguous names come back as candidates, never a guess.

Or, without an agent: pctx remember "Amina Hassan" "prefers short emails" --org "Open City Lab" and pctx brief "Amina Hassan". Five worked scenarios live in docs/use-cases.

What it remembers, and what it never does

How it compares

The dated, sourced version with vendor documentation links is in docs/privacy-and-safety.md.

Security model

This project executes local Python with the launching user's filesystem permissions. Ordinary MCP discovery excludes elevated sensitive context and full export. Operator-gated tools require process environment flags; models cannot enable them through arguments. Vault export is intentionally CLI-only.

The database is plaintext SQLite by default. On Unix-like systems a new one is created 0600, so other local accounts cannot read it. That is a boundary between accounts, not encryption, so pair it with full-disk encryption or opt into SQLCipher at-rest encryption (uv sync --extra encrypted, key read only from PEOPLE_CONTEXT_DB_KEY). See database file permissions and optional at-rest encryption.

Going further

• Loopback HTTP for clients that cannot spawn stdio: people-context-mcp --http --host 127.0.0.1 --port 8765. Unauthenticated and local-only by design; prefer stdio. See docs/cli.md.
• Semantic search: uv sync --extra semantic && pctx reindex --semantic downloads a pinned multilingual Model2Vec model once; server startup and search stay cache-only.
• Obsidian: pctx export-vault --output ~/PeopleVault writes a deterministic, browsable vault, and a read-only Obsidian plugin renders live briefs. See docs/obsidian-plugin.md.
• Import: pctx import stage SOURCE PATH then review and commit, over email, mbox, vCard, .ics, LinkedIn, Outlook, and WhatsApp exports. Agents can stage extracted candidates the same way. See docs/import.md.
• Reports and maintenance: pctx stale, pctx upcoming, pctx timeline, pctx doctor, pctx stats.
• Backup and second device: pctx sync push --output DIR and pctx sync pull --input PATH.
• Docker: docker run --rm -i -v people-context-data:/data ghcr.io/jinyangwang27/people-context:latest. A convenience image, not a sandbox. See docs/docker.md.
• Database location: --db, then PEOPLE_CONTEXT_DB, then the XDG config file, then the shared ~/.pctx/people.db. Inspect with pctx db-path -v.

The full command reference is in docs/cli.md; the MCP tool inventory and response contracts are in docs/mcp-interface.md; what stays stable across releases is in docs/compatibility.md.

Architecture

The codebase follows ports and adapters:

adapters (SQLite, MCP, filesystem, imports, CLI)
        ↓ implement
ports (narrow Protocols)
        ↑ used by
app (use cases and policy)
        ↓ operates on
domain (entities and values)

Dependencies point inward. Vocabulary normalization and graph caps live in app/domain; recursive SQL and file writing live in adapters. One composition root wires both stdio and HTTP. See docs/architecture.md.

Documentation

Contributing

Issues and pull requests are welcome; see CONTRIBUTING.md for the architecture rules, validation commands, and a list of good first issues. Questions and show-and-tell go to Discussions.

If people-context is useful to you, a star helps other people find it.

License

MIT. See LICENSE.