Document Carabase — Whitepaper & Product Brief
Version v0.4 · 2026
Status Early access
Sections §01 — §06
§00 · Cover Early access · 2026

The engine that remembers last Tuesday.

A personal context engine on your hardware.

Carabase reads your calendars, mail, code, meetings, and notes through connectors you control. Ask any question; the engine returns the answer in under 10 ms with sourced citations. Bodies stay on your hardware. Plugs into Claude, OpenClaw, or any MCP-capable agent.

Carabase · Summon
⌘K
>what did sarah commit to last tuesday on the term sheet
Answer

Sarah confirmed the indemnity cap stays at 1× ARR. She asked for the survival period to move from 12 to 18 months — you said you’d come back on that by Friday.

gmail sarah@acme · re: term sheet v3 tue 04:32 pm
granola acme — weekly check-in tue 03:15 pm
note “term sheet thoughts” tue 05:01 pm
Carabase resolves the question across four substrates in under 10 ms and returns an answer with three cited sources.
Abstract §00.1

Carabase is a personal context engine on top of a personal data mesh. The mesh is your calendars, mail, code, browser, photos, messages, and health — pulled in through connectors you control, stored on hardware you own, with retention and materialization you set. The engine is what runs over it: a typed knowledge graph with provenance, a memory network for sense-making, and a deterministic router that returns in under 10 ms.

Both are yours, materially. This document lays out the problem the category got wrong, the architecture we built instead, and the surfaces — Host, Desktop, CLI, iOS, MCP — through which you drive it.

Retrieval
< 10 ms deterministic router
Substrate
Postgres + pgvector, on your hardware
Surfaces
Host · Desktop · CLI · iOS · MCP
§01 · How it knows trust block

Your data mesh. Your hardware. Your call on every connector.

Carabase reads from sources you authorize — through connectors with explicit scopes, OAuth or local-file access, and a visible audit trail. Bodies stay on your hardware: the substrate is Postgres + pgvector on a machine you own; the master key lives in your env; connector tokens are encrypted at rest. LLM inference is the part that travels — point it at ollama for fully local, or pick an API per role.

Connect
15+ sourcesGoogle, GitHub, Granola, Beeper, Apple Notes/Mail/Photos, Safari, Strava, Spotify, Readwise, iOS HealthKit. See the matrix →
Materialize
4 tiersFull body / hollow (TL;DR + entity graph) / metadata-only / dark. Per-rule, per-account.
Retain
Tightest cap winsWorkspace default · per-account cap · per-rule cap. Auto-purge; once purged, no silent re-fetch.
Review
HITL queueLow-confidence inferences land in Review. You accept, reject, or silence — nothing speculative escapes by default.
Gmail Granola GitHub Apple Notes Safari Beeper Substrate Postgres + pgvector your hardware MCP · Agents

The full architecture lives at /architecture. The connector matrix lives at /connectors.

§02 Surfaces p. 02

§02 One substrate. Four surfaces.

Four surfaces, all yours, all talking to each other over a private mesh that never touches the public internet — plus a fifth row for agents.

§02.1 · Host
The engine

Postgres + pgvector + typed knowledge graph + memory network + MCP server. Runs on your own hardware. Zero public ports. Nightly harvest, synthesis, dream cycle.

Self-hosted · Open source
§02.2 · Desktop
The interface

Rust. Tauri. Native. Cmd+K everywhere. Daily log, summon panel, ambient graph, command palette, ingestion controller, HITL review queue.

macOS · Linux
§02.3 · CLI
The terminal

carabase init in 7 steps. Pipe a note from stdin. Stream an agent run with --watch. Search semantic + hybrid. Wire it into your shell. Headless-first.

carabase · cb alias
§02.4 · iOS
The capture

Share extension. Siri Shortcuts. Lock Screen widget. HealthKit + Photos + Significant Locations + Shortcuts inbox. Walking-with-a-podcast in < 3 s.

iOS
§02.5 · MCP + SDK
For agents and integrations

Every retrieval and write surface is exposed as an MCP tool. Claude Desktop, OpenClaw, your custom agent — they all query the same substrate. Plugins for new connectors. Open-source Host. Heterogeneous stacks welcome.

MCP · Integration SDK · Open
§03 Speed p. 03

§03Speed is the first architectural choice.

<10
MILLISECOND RETRIEVAL

Median p50 across 1,000 internal-corpus queries on an Apple M2 Pro Mac mini, hybrid + graph + temporal paths. Excludes any LLM call. Methodology: /architecture.

No LLM in the retrieval loop.

Every other "AI" search pipeline: query → embed → vector search → feed results to LLM → LLM picks the best one → return. Slow because an LLM is involved. Also a pyromaniac's approach to tokens.

Ours: a deterministic router across six substrates — a typed knowledge graph for relationships, a semantic index for meaning, a metadata index for time and source, an entity resolver for names, a hypothesis verifier for claims, and an associative memory network for sense-making. Postgres, pgvector, GIN indexes, and a graph traversal each do their job in single-digit milliseconds. We let them.

§04 What the Engine Does p. 04

§04Seven properties. None suck.

§04.1 · Temporal
Knows when something became true — and when it stopped. "Who owned this in March?" is a different query from "who owns it now." Both answerable.
§04.2 · Provenance
Every edge cites a source and a confidence score. Verify a claim. Get citations. Chatbots invent citations — we show you the real ones.
§04.3 · Layered ingestion
Four materialization tiers on every artifact — full body, hollow (TL;DR + entity graph), metadata-only, dark. Per-rule, per-account, per-source. You set the defaults; the engine respects the cap. Bodies you've purged stay purged — by default, Carabase will not re-fetch from upstream even if it could.
§04.4 · Agent-native & open
MCP server. Integration SDK. Claude Desktop, OpenClaw, whatever ships next — they all query your substrate. Heterogeneous stacks, welcome.
§04.5 · Yours, materially
Zero public ports. AES-256-GCM at rest. Row-level security. Open source Host. If we disappear tomorrow, your engine keeps running.
§04.6 · Moss & ember
Every entry shows you if you authored it (moss) or the engine harvested it (ember). An epistemic contract — not a design flourish.
§04.7 · Human-in-the-loop curation
When the engine infers a relationship at low confidence — "User → works_with → Sarah from Mailchimp" — it doesn't quietly promote it. The inference lands in a Review queue. You accept (it becomes canonical), reject (it's soft-deleted), or silence (don't ask again about this pair). Optional: dispatch an in-loop agent that does web search and stamps a verdict on the card before you see it. Auto-accept threshold configurable. Nothing speculative escapes review by default.
§FAQ Frequently Asked appendix A

§FAQAnswers, before you ask.

The questions that come up most often. Longer answers live in the manifesto.

What is a personal context engine?
A persistent, temporal, queryable model of everything you know, who you know it from, when you learned it, and how much you trust it. It runs on your hardware. It is controlled by you. Any AI tool can plug into it via the Model Context Protocol.
Is Carabase really self-hosted?
Yes. The Host service — Postgres + pgvector + knowledge graph + MCP server — runs on hardware you own. The Desktop and iOS clients connect over a private Tailscale mesh that never touches the public internet. Bodies stay on your hardware.
When does early access open?
Rolling invites through 2026, onboarding 50 engineers at a time. Join the waitlist to get in the queue.
Is there a CLI?
Yes. The carabase command (alias cb) is a peer to the Desktop UI, not an afterthought. cb init is a 7-step interactive wizard. cb agent run "summarize my week" --watch streams a live agent run. cb chat ask "what did I commit to last Tuesday" opens an SSE chat session. cb import upload notes.zip migrates from Obsidian / Logseq / Notion / Bear / Roam / Reflect. Every write verb takes --idempotency-key so retries are safe. cb --json for scripting, cb --watch for streaming, cb --env to switch between dev / staging / prod. See the CLI page.
You said “capture everything.” How is that not surveillance?
It's surveillance against you only if someone else holds the keys. With Carabase the substrate runs on your hardware, the master key is in your env, the network is your Tailscale mesh, and the connector tokens are encrypted at rest. Above that, three knobs make breadth livable: (1) materialization tiers limit which bodies actually land — most edge-source items default to metadata-only with bodies fetched on demand; (2) retention caps (workspace / per-account / per-rule, tightest wins) auto-purge bodies on a schedule; (3) HITL review keeps speculative inferences out of canonical until you accept them.
What runs in the cloud and what doesn't?
Your data substrate — Postgres, pgvector, the knowledge graph, the memory network, every artifact body — runs on your hardware. LLM inference is the part that can run anywhere you point it. Pick Ollama and everything is local. Pick OpenRouter, Anthropic, or OpenAI and your prompts (not your full substrate, just what you ask about) travel through their API. Carabase routes per role: utilityHigh for reasoning, utilityLow for high-volume background work, embedding for vectorization, visualUtility for vision, image_embedding for visual similarity. Each role picks a provider independently; mix and match.
§06 · EARLY ACCESS

Imagine the answer arriving in the same breath as the question.

Carabase is in early access for technical users — engineers, builders, executives, and anyone whose bottleneck is context. Join the waitlist.

Set in Instrument Serif, Geist, IBM Plex Mono Compiled 2026 · v0.4 Hosted on hardware you own