MyClip helps you remember what you were working on. It captures the focused window or its display on your Mac and uses Codex or Claude to turn screenshots into searchable notes, connected knowledge, and suggested tasks.
English · 简体中文 · Español · Français · Deutsch · 日本語 · 한국어
45-second HTML demo · Sample data · Chinese interface.
- Revisit your work. Browse a screenshot timeline and check the sources behind your notes. Every screenshot has a local OCR text document you can view, copy, or open.
- Build a personal knowledge library. Search, edit, and link notes across projects, topics, and daily work. Your notes are Markdown files you can open in other editors.
- Keep track of next steps. Review suggested tasks, confirm what matters, and follow progress on a task board. Switch between Kanban and Reports to review daily, weekly, or monthly work reports grouped by project and progress.
- Give your AI tools context. Let Codex, Claude Code, Claude Desktop, Cursor, or OpenCode search your saved memories.
Requires macOS 13 or later and a Codex or Claude account. Installing the agent connector also requires Node.js 22 or later. Download the Apple Silicon or Intel DMG from Releases. These community builds are self-signed and not notarized; see opening the app and permission recovery. Local build instructions are below.
- Open MyClip and choose a default Agent in onboarding. MyClip detects available local agents and saves your selection after a successful connection. You can change it later in Backstage; multiple agents can be connected, but only one is enabled at a time.
- Grant Screen Recording and Accessibility permissions. Capture starts automatically whenever MyClip is open, including after permissions are granted.
- Work as usual. By default, MyClip captures the focused window after a click within five seconds of mouse movement, after three seconds without scrolling following a downward scroll, or after a letter key followed by Return. Automatic organization turns new screenshots into Memory.
- Explore Memory for notes, Timeline for screenshots, and Kanban for suggested work.
To use your memory in an AI tool, enable MyClip MCP in Settings, select your clients, and apply the configuration. Restart the client or start a new session afterward.
Claude Code (CLI) handles background screenshot organization through ACP. Claude Desktop has a separate entry for opening the app and configuring MCP memory access in Chat and local Code sessions; it cannot be enabled for background organization. Desktop configuration is saved to ~/Library/Application Support/Claude/claude_desktop_config.json, preserving existing servers and leaving the CLI configuration untouched. Fully quit and reopen Claude Desktop after configuring it.
MyClip extracts Chinese and English text on-device with Apple Vision, independently of AI organization. In a screenshot's detail view, select OCR to read or copy its text, or choose 打开文档 to open the UTF-8 .txt file saved beside the original image. Repeated screenshots share one image and text document. Screenshots without readable text receive an empty document; recognition failures can be retried.
Existing screenshots are processed in the background after startup. OCR documents and their search index expire with the original images according to the retention setting. Saved Memory notes remain. The app targets macOS 13 and later, including macOS 27; macOS 13 uses a single-frame ScreenCaptureKit stream with the same application exclusions as newer systems.
Screenshots enter a persistent waiting pool as soon as they are saved. Automatic organization waits five minutes from the oldest pending capture, then sends up to eight original screenshots in chronological order to the enabled Agent. Every capture trigger uses the image. OCR remains a local text document and search source; it is not sent as an organization input. New captures do not reset the timer. Only one batch runs at a time, with at least five minutes between batch starts.
The Backstage page shows the waiting count, countdown, and current batch. Organize Now starts one batch early. A failure or interrupted run pauses automatic processing until you retry or resume; screenshots and existing Memory files are retained. A batch keeps its screenshot sources across retries and restarts. 按图片重新整理 in screenshot details sends the original image again. Enabling another agent reassigns captures waiting for automatic organization. Running batches finish with their original agent. Existing jobs and retries also keep their original agent and wait until it is enabled again.
Each batch uses an independent temporary session. Claude receives persistSession: false; MyClip's Codex app-server proxy enforces ephemeral: true and rejects a backend that does not acknowledge it. The agent process closes after each run. Old saved conversations are neither resumed nor deleted. The next batch receives the fixed organization rules, only the previous successful batch's handoff (at most 4 KiB), current timestamped inputs with their source IDs, app/window and trigger metadata, and existing task context. Related Memory files are read on demand. The handoff records saved file changes and source IDs, excludes conversation history and Memory bodies, and is replaced only after Memory publishing succeeds. Failed retries start fresh and inspect the current files.
Without an enabled agent, screenshots remain in the waiting pool. Disabling an agent stops new tasks while allowing the current task to finish. MyClip remembers the enabled agent and reconnects it at startup. Older default-agent preferences do not enable an agent automatically; enable one explicitly after upgrading.
MyClip starts every organization and task-discovery session with Full access: agent-full-access for Codex and bypassPermissions for Claude Code. File access, edits, commands, network access, and MCP tool calls proceed without per-operation confirmation cards. Any remaining tool permission requests are handled automatically for the active session, while cancelled tasks reject late requests. Tool activity remains available in the execution record.
The Backstage page tracks reported token consumption across organization, task discovery, and retries, with totals per agent and per batch. Usage is stored locally when each request ends; older or unreported usage is shown as unavailable, not zero. Context-window occupancy is not counted as consumption. Active batches show their current stage, elapsed time, and time since the last progress update. Claude ACP uses Claude Code's existing login and network settings, so a configured local proxy must be running.
Organization stops after five minutes without new thinking, response text, tool activity, or permission activity in the current session. A batch can continue while it makes progress, up to a fifteen-minute total limit. Usage-only updates and other sessions do not extend the timeout.
Memory keeps screenshot observation time (observed_at) separate from file update time (updated_at). Now shows when its evidence was captured; older evidence cannot replace a newer Now page. Missing observation time remains unknown. The organizer consolidates event history in Daily, keeps conclusions in project/topic pages, and moves resolved items out of Inbox. Repeated or incidental browsing does not require a new permanent note.
For newly organized pages, source_ids contains actual screenshot IDs cited in the text. The batch's reference screenshots are retained separately in context_source_ids; they are not evidence for every statement. MCP search results carry the cited source IDs of their paragraphs, and memory_get summarizes the screenshots behind a page (count, time span, apps). Existing notes remain readable and adopt these rules when organized again; notes are not bulk-rewritten during upgrade.
Memory search in the app and MCP uses the same SQLite FTS5 ranking. Keywords match any term; exact titles and declared aliases come first, then a mix of each note's best-matching paragraphs and whole-note BM25, then file modification time. Scoring by paragraphs keeps long notes from outranking the paragraph that answers the question. Literal substring matches remain available for Chinese text and punctuation. An empty query still lists recently edited notes.
MCP offers two read-only tools. memory_search takes query (the question or keywords) plus optional since, until, app and limit (default 10), and returns ranked results with path, title, a short snippet (up to two matching paragraphs of at most 300 characters each, with Wikilinks shown as their labels and citation IDs left out), links (the pages the shown lines point to, as path or path#heading), time, the paragraphs' cited sourceIDs and source apps. memory_get takes a path from those results, optionally with #heading to read one section, and from/lines for line ranges; it returns the Markdown lines, the page's links grouped by heading, its backlinks newest first (each with the line that holds the link and its date), and a summary of the screenshots behind the page. Unknown arguments are rejected rather than ignored.
Search follows Wikilinks. The strongest matches, plus both ends of links whose line matches the question, seed a two-step spread over the link graph: one hop from every seed and a second hop only through entity pages (a Daily note → a person or topic page → another Daily note). Each link is weighted by how well its line matches the question and damped by how many links the page it enters has, so hub pages such as Now.md do not flood the results; root files and Wiki/Archives are left out. Pages reached this way are ranked together with direct matches and carry via: one or two hops, each with the page that holds the link, its heading, the line itself and its date. Links indicate association, not proof of a factual relationship.
time and the since/until filters describe when the content happened: an annotated event on a paragraph, else a cited screenshot, else the file edit time. since is inclusive and until exclusive; both accept ISO 8601 timestamps. With app, notes with events match through an event paragraph that also cites a screenshot from that app, and notes without events need one cited screenshot that satisfies both the range and the app. Unknown event dates are never replaced by capture or edit dates. Dated pages without a range gently favor recent days.
Event annotations are stored immediately before their paragraph, without a blank line, as an HTML comment. They survive Markdown copies and index rebuilds. For example:
<!-- myclip-event {"start":"2026-09-10T00:00:00+08:00","end":"2026-09-11T00:00:00+08:00","precision":"day","evidence":"2026年9月10日"} -->
2026年9月10日，客户会议结束。来源：截图 `REPLACE_WITH_ACTUAL_SOURCE_UUID`。Use actual evidence and a real cited screenshot ID. Supported precision values are day (a local calendar day), range (an explicit interval, end exclusive), and instant (end omitted or equal to start). Timestamps require an explicit UTC offset. Search reports the event start as the result's time; this records the note's assertion, not independent verification of the source. Invalid dates, conflicting annotations, annotations in code examples, missing paragraph citations, or an evidence expression absent from the paragraph do not produce indexed event times. Relative dates require a known original-message time and timezone; the organizer must retain the expression and explain its conversion. Capture time and edit time never supply a missing event date. Existing notes remain searchable without annotations and receive them when relevant evidence is organized; upgrading does not invent or rewrite their dates.
Passage, event and link indexes are derived SQLite data. They are refreshed on edits, deleted with their note, and rebuilt for older libraries without changing Markdown. For questions that need more than two hops, an agent reads a page with memory_get and follows the links or backlinks it lists.
Temporary sessions cannot be reopened in Codex or Claude. Inspect their execution details in MyClip instead; each record also shows its image/text input counts. These settings prevent resumable local agent conversations, not the model provider's service-side data retention.
Click an organization record in Backstage to inspect each request, including retries: tool calls, command arguments, results, file locations and edits, agent replies, input/output tokens, cache reads/writes, and reported cost. Tool records are saved while the request runs and remain available after cancellation or restart. Cost uses the difference between reported cumulative session amounts; missing reports or an unknown session baseline remain unknown. Older records retain their existing token totals but cannot recover tool details that were never saved.
- Choose what to capture. Settings groups capture into three controls: scope (focused window by default, or the display containing it), independent mouse triggers (click within 5 seconds of mouse movement and pause for 3 seconds after scrolling down), and keyboard trigger (letters then Return by default, or every Return). Capture runs automatically while MyClip is open; quit the app to stop it. You can exclude specific apps; full-display capture also filters excluded apps.
- Keep your library locally. Screenshots and notes are stored on your Mac. Original screenshots expire after 30 days by default; saved notes remain. You can change the retention period in Settings.
- Decide when to use AI. Organization uses your selected agent's model service, which may process screenshots and notes in the cloud. Turn off automatic organization to keep new captures local until you choose to process them.
Build from source
Requires Xcode 26, Swift 6.2, XcodeGen, Python 3, and Node.js. See the development guide for source layout, prompts and test groups.
xcodegen generate
bash Scripts/test.sh
xcodebuild -project MyClip.xcodeproj -scheme MyClip -configuration Debug buildTo package a DMG, run Scripts/package_dmg.sh. The output is dist/MyClip-<version>.dmg.
For separate Apple Silicon and Intel packages, run MYCLIP_ARCH=arm64 bash Scripts/package_dmg.sh or MYCLIP_ARCH=x86_64 bash Scripts/package_dmg.sh. Their filenames end in -arm64.dmg and -x86_64.dmg respectively. After the one-time community signing setup, pushing a v<version> tag triggers GitHub Actions to test and package both architectures, then publish a Release with both DMGs, the public signing certificate, and SHA256SUMS. The tag must match CFBundleShortVersionString, with release notes in docs/releases/v<version>.md.