Continuum Core

ci

Your whole desktop context — everything you've seen, read, typed, and heard — as a single MCP your agent can use to understand you and how you work.

Every agent starts blank, so you become the courier — re-pasting and re-explaining across a dozen apps a day. Continuum remembers what happens across your apps, on-device and private, and serves it to any agent over MCP. Ask Claude "what was I just doing in X?" and it already knows.

▶ Watch the demo

Get started

Hand it to your coding agent — paste this into Claude Code, Codex, or Cursor:

Set up Continuum so you can see my desktop context: install it (npm i -g continuum-core), start capture in the background (continuum start & — I'll grant Screen Recording when macOS prompts), then add the Continuum MCP server from continuum mcp-config. From then on, use its recall, catch_up, and profile tools on your own judgment to understand what I'm building, how I think, and my taste — not only when I explicitly ask about my activity.

Then just ask it "what was I doing in X?" — that's the whole setup.

Prefer by hand? (needs Node 18+)

npm i -g continuum-core
continuum verify       # prove it works in 30s — no keys, no setup
continuum start        # live capture (grant Screen Recording once)
continuum dashboard    # your timeline + search at localhost:3939
continuum mcp-install  # connect it to Claude Desktop, then restart Claude

What you get

• Sees your screen — on-device OCR of the focused window (deduped to content, not noise).
• Cited & dated — every recalled moment comes back with its source app, owner, and an unambiguous calendar date ("as of Jun 18, 2026") plus a citation you can open — so your agent attributes what it says and never passes off a stale fact as current.
• Understands who you are — an out-of-band dreaming pass consolidates the raw firehose into small, cited memory files (who you are · projects · people · taste · decisions) your agent reads to genuinely understand you, not just recall moments (continuum dream · continuum memory).
• Learns how you work — standing preferences like "be concise" or "run the tests before the PR" apply to every agent automatically; you stay in control (dashboard or continuum preferences).
• Local-first & read-only to agents — your memory lives in ~/.continuum and the agent only reads it (a prompt-injected agent can't poison it). Capture, search, and indexing run on-device; only the snippets your question needs — and the scheduled dream digest — go to the model you choose, through a single audited egress (an append-only ledger of exactly what left the machine, enforced by a build-failing check) and best-effort secret/PII redaction. The store can be encrypted at rest (opt-in). We don't claim nothing leaves; we show you exactly what does.

Free & local: capture, retrieval (hybrid lexical + local embeddings + RRF), and preferences run on-device for free. Deep memory (dreaming) needs a capable model — a strong local instruct model, or an OpenAI / Anthropic key.

How it works

Two tiers over one episode store. Continuum captures your screen and groups it into episodes. The live tier indexes them so any agent can answer "what was I doing?" over MCP (recall / catch_up) — local and free. An out-of-band dreaming pass reads the same episodes and consolidates them into the durable memory files your agent reads to understand you (profile): verify · organize · enrich, grounded in the source moments. The LLM never touches the capture path. The stages are importable modules (a useful tool is ~20 lines, see examples/). Deep dive: architecture.

Develop

From a clone of the repo:

npm test                                                   # full suite, no network
swiftc daemon/stage1/screen.swift -o daemon/stage1/screen  # build the capture helper

continuum eval reports capture-quality metrics over local fixtures. Contributions under DCO (git commit -s). License: Apache-2.0.