Persome

The local-first Personal Model Runtime for macOS. Persome observes the apps
you already use, turns cross-app activity into an inspectable model of a real
person, and serves that model to Chat and MCP agents.

Star Persome on GitHub to
follow the Runtime and help prioritize the next MCP integrations.

Actual /model screenshot produced by scripts/sample_demo.py --showcase: 424
synthetic Points, 146 Lines, 12 Faces, 4 Volumes, and 1 Root. It contains no
personal data.

Product job

Persome runs quietly on one Mac and does four jobs:

1. Collect focused macOS Accessibility (AX) context across apps, with an
   optional on-device OCR fallback for AX-poor surfaces.
2. Model observations into sourced facts, evolving relations, stable
   patterns, cross-domain structure, and one current Root.
3. Serve local memory and model tools over MCP, plus an optional terminal
   Chat that uses the same tools.
4. Give control back through receipts, time travel, correction, export, and
   deletion.

This is the Runtime, not a hosted account or a single assistant's private
memory. One local model can be used by Claude Code, Codex, Cursor, or another
trusted MCP client.

Five-minute sample demo

See the whole model without an API key, Accessibility permission, or access to
your real ~/.persome data. This path requires Git and
uv:

git clone https://github.com/Persome-ai/persome-core.git
cd persome-core
uv run python scripts/sample_demo.py

Add --showcase to render the denser, still fully synthetic model used in the
README image.

The script opens http://127.0.0.1:8743/model, serves MCP at
http://127.0.0.1:8743/mcp, and deletes its temporary synthetic data when you
press Ctrl-C. To inspect the exact search, receipt, and snapshot payloads:

PERSOME_LLM_MOCK=1 uv run python scripts/sample_demo.py --json

With the sample server still running, verify the actual MCP transport from a
second terminal:

uv run python scripts/verify_sample_mcp.py

This sample path is deliberately separate from the real-data path below.

Quick start with your data

Requirements: macOS 13 or newer, Xcode Command Line Tools, and a Python build
with SQLite 3.42+ (the installer verifies the secure FTS capability). The installer
finds or installs uv, provisions Python 3.11-3.13, compiles the Swift AX
helpers, generates the local screenshot-encryption key, enables and verifies
local OCR, requests Screen Recording, and offers to register detected MCP
clients. Its fallback uv download is version-pinned and checked
against repository-pinned SHA-256 digests; the Runtime environment is installed
from the committed uv.lock, and the complete build-backend closure is
hash-constrained rather than resolved afresh.

git clone https://github.com/Persome-ai/persome-core.git
cd persome-core
bash install.sh

persome doctor
persome ocr status --check
persome start
persome model open

Grant Accessibility to the terminal or app that launches Persome in
System Settings -> Privacy & Security -> Accessibility. This permission is
required to read focused AX text and structure. The installer enables bundled
local OCR, requests Screen Recording, verifies the isolated OCR worker, and
opens the correct settings pane when permission remains denied. OCR supplies
text for AX-poor apps such as WeChat and Feishu; pixels never enter an LLM
prompt. Persome does not require Full Disk Access.

# Recheck or repair OCR onboarding; disable is always explicit and reversible.
persome ocr setup
persome ocr status --check
persome ocr disable

An LLM is optional for collection and BM25 recall, but required for semantic
modeling. During installation, the provider wizard asks you to choose a service
and enter its API key. Persome supplies that provider's endpoint and default
model, tests completion and tool calling, and only then saves the route. Existing
keys are detected automatically. API keys go to the owner-only
~/.persome/env file under the provider-neutral PERSOME_LLM_API_KEY name;
provider-specific environment variables are import sources only. The non-secret
route goes to ~/.persome/config.toml. Nothing ships with a key.

# If provider setup was skipped during installation:
persome llm providers
persome llm setup
persome llm status --check

# Restart after changing the active provider:
persome stop || true
persome start

Persome speaks two wire protocols: native Anthropic Messages and
OpenAI-compatible Chat Completions. Presets cover Anthropic, OpenAI, DeepSeek,
OpenRouter, Gemini, Groq, Mistral, xAI, Qwen, Moonshot/Kimi, Zhipu GLM,
SiliconFlow, Together, Fireworks, Cerebras, Azure OpenAI, Ollama, LM Studio, and
vLLM. custom-openai and custom-anthropic accept another compatible endpoint.
Azure and custom endpoints use a clearly marked advanced setup path. A preset
means the route is configured, not that every model has the necessary
capabilities; Persome warns when the default model cannot call tools.

Active work is reduced every five minutes by default. A first useful recall is
therefore expected within ten minutes of valid capture plus a working semantic
provider; persome status, persome model status, and the viewer explain sparse
or degraded states instead of inventing geometry.

Proof points

Local-first

• Durable Markdown, SQLite/FTS5, model snapshots, and logs live under
  ~/.persome unless PERSOME_ROOT is set.
• AX is the default signal. Optional PP-OCRv6 runs locally in an isolated
  subprocess with bundled weights.
• The HTTP/MCP server is restricted to loopback (127.0.0.1 by default), requires an owner-local
  bearer on API/MCP routes (or its one-use derived viewer capability), and emits no telemetry.
• Only configured semantic stages send derived text to the selected provider's
  LLM or embedding endpoint.

Cross-app

The Swift watcher reads the focused AX tree across native and browser apps.
Persome normalizes focused element, visible text, window, application, URL, and
time into one capture and session pipeline. OCR is a fallback, not a parallel
cloud recorder.

Agent-ready

• Authenticated streamable HTTP MCP: http://127.0.0.1:8742/mcp
• stdio MCP: persome mcp
• Local Chat: persome chat
• Stable model contract: persome model export and GET /model/graph
• Evidence tools: search, read_receipt, verify_fact, and
  get_model_snapshot

Connect an MCP client

Register an owner-local stdio server. These clients launch it on demand, so the
daemon does not need to be running and no bearer is copied into their config:

persome install claude-code
persome install codex
persome install claude-desktop
persome install opencode

# Generate a stdio config that can be merged into Cursor's MCP config:
persome install mcp-json --filename persome-mcp.json

The canonical JSON shape is:

{
  "mcpServers": {
    "persome": {
      "command": "persome",
      "args": ["mcp"]
    }
  }
}

See MCP client setup and verification for authenticated
HTTP configs, uninstall commands, and privacy boundaries.

Real MCP query with a cited answer

The following result is generated by the committed synthetic sample through the
same search and read_receipt implementation exposed by MCP.

Tool: search
Input: {"query":"When does the user prefer focused writing?","top_k":2}

Top result:
  id:        20260701-0800-d4e5f6
  path:      project-work.md
  timestamp: 2026-07-01T08:00
  content:   The user reserves mornings for focused writing and review.

Tool: read_receipt
Input: {"entry_id":"20260701-0800-d4e5f6"}

A grounded client response can then say:

The user prefers mornings for focused writing and review.
[project-work.md, 2026-07-01 08:00;
receipt 20260701-0800-d4e5f6]

The receipt is resolvable, the superseded earlier statement remains available
as history, and the answer does not rely on the model's unsupported memory.

Benchmark and verification status

This repository reports Runtime engineering evidence, not a paper-quality
personalization benchmark.

The sample uses synthetic fixtures and cannot establish recall quality on a
real person. No cross-user benchmark, next-action accuracy, latency percentile,
or comparison win is claimed. The launch machine's three isolated source
installs had an 11.896-second median with a warm uv cache; conditions and
limitations are recorded in benchmark scope.

Why Persome

These projects solve adjacent but different jobs:

Persome is not a replacement for a full screen archive, a hosted vector memory,
or a provider's preference feature. Choose it when the core requirement is a
local, cross-app, auditable model that multiple agents can query.

How it works

flowchart LR
  AX[macOS AX watcher] --> S0[S0 debounce]
  OCR[Optional local OCR] --> S1[S1 normalized capture]
  S0 --> S1
  S1 --> BUF[Capture buffer]
  BUF --> TL[1-minute timeline]
  TL --> SES[Deterministic sessions]
  SES --> DELTA[5-minute memory delta]
  DELTA --> PL[Points and Lines]
  PL --> FV[Faces and Volumes]
  FV --> ROOT[Root]
  PL --> RET[BM25 and optional dense retrieval]
  FV --> MCP[MCP, Chat, export, viewer]
  ROOT --> MCP
  RET --> MCP

Every modeled object keeps source receipts and bitemporal history. A sparse
store can truthfully contain Points and Lines without a Face, Volume, or Root.
The viewer shows that incomplete state rather than fabricating one.

Read Runtime architecture, the
model contract, and the detailed
maintainer architecture.

Inspect, correct, export, and delete

# Inspect
persome status
persome model status
persome faces-report
persome contradictions
persome model open

# Correct or revoke one memory while retaining its audit trail
persome correct --help
# Agents can also call MCP correct_memory.

# Export a redacted owner-only snapshot (0600)
persome model export

# Delete model memory, or all captures/timeline/model state
persome stop
persome clean memory
persome clean all

For a complete uninstall that preserves personal data by default:

bash uninstall.sh

# Explicitly remove the remaining data, config, env, exports, and logs:
bash uninstall.sh --delete-data --yes

Client registrations are removed separately and idempotently:

persome uninstall claude-code
persome uninstall codex
persome uninstall claude-desktop
persome uninstall opencode

See operations and data control for exact paths, backup
advice, export sensitivity, reset behavior, and manual removal steps.

Privacy boundary

• Personal data remains local until a configured model stage or connected agent
  sends selected text to its own provider.
• MCP capture tools can return raw screen text, titles, URLs, and focused-field
  values. Bearer/stdio access is a personal-data capability; connect only
  clients you trust.
• Model-generated memory never becomes trusted Chat skill instructions;
  unsafe/external Chat tools require exact one-shot terminal approval.
• Screenshots are omitted from MCP by default and encrypted at rest when
  retention is enabled.
• persome model export is redacted by default; --raw is an explicit opt-out.
• There is no built-in remote account, sync service, telemetry, meeting audio
  capture, computer-use actuation, or filesystem profiler.

Read Security and privacy before using real personal
data, and report vulnerabilities through SECURITY.md.

Platform support

Python 3.11-3.13 with SQLite 3.42+ is supported by the installer. See
operations and troubleshooting.

Persome and Personome

Persome is this open-source Runtime and project name. Personome is the
research term for the learned model of one person: a dynamic state assembled
from sourced observations, relations, stable patterns, and higher-level
structure. The product name stays Persome in commands, packages, paths, APIs,
and documentation.

Paper and architecture-note status

This repository ships the executable Runtime and an implementation-oriented
architecture note. The architecture documents are not a peer-reviewed paper,
and the Runtime's synthetic gates are not publication benchmarks. The paper,
benchmark suite, data statements, and project publication will live as separate
artifacts with independent licenses before release. See
licensing boundaries and benchmark limitations.

Roadmap

The public roadmap is issue-driven:

• more tested MCP client integrations;
• richer first-run permission diagnostics;
• explicit import/export interoperability;
• Intel and future-macOS compatibility evidence;
• a separate, reproducible personal-model benchmark suite.

Browse starter issues or
start a design question in
Discussions.

Contributing and community

Read CONTRIBUTING.md, follow the
Code of Conduct, and use SUPPORT.md to choose
the right channel. Every commit requires DCO sign-off, and CI blocks known
secrets, personal data, non-English source text, contract drift, lint failures,
and offline regressions. Third-party Actions are pinned to reviewed commit SHAs
and workflow permissions default to read-only.

Support Persome

If an inspectable, user-owned personal model is useful to your agents,
star Persome on GitHub and
share the MCP client or workflow you want supported in
Discussions.

License

Runtime code is Apache-2.0. Paper, benchmark, project-note, third-party, and
personal-data boundaries are explained in LICENSES.md. Required
incorporated-work notices remain in NOTICE and
THIRD_PARTY_NOTICES.