Pieces

Pieces acts in one outside service today. Once a user connects Google Calendar through Google's own sign in, conversational search can create, update, reschedule and cancel events and invite attendees, and the Meeting Prep summary adds prep blocks before upcoming meetings. GitHub, Slack and other connectors are listed as coming soon, and there is no way to add a connector of one's own. Everything else the assistant reaches is read only: Perplexity web search with citations, local files, and browser history and bookmarks.

Sourcedocs.pieces.app/products/desktop/connectorsread 2026-11-09

Each chat or summary runs as one agent working over several turns. When a first search of memory leaves a gap, the agent goes back for more, cross references summaries, events and people, and can combine memory search, web search, calendar actions and local files in one answer; a reflection mode has it check its own reasoning. Summaries can also be scheduled to run on their own. There is no workflow builder, branching logic or second agent to hand work to.

Sourcedocs.pieces.app/products/core-dependencies/pieces-os/long-term-memoryread 2026-11-09

Answers draw on sources gathered for each question. The assistant can search the user's local files by path, grep their contents and read them, run Perplexity web searches with citations, look up browser history and bookmarks, and pull upcoming meetings from a connected Google Calendar. Each of these sources is read when a question needs it rather than indexed in advance, and no separate knowledge base holds a team's documents, wikis or tickets for retrieval.

Sourcedocs.pieces.app/products/core-dependencies/pieces-os/long-term-memoryread 2026-11-09

Before the assistant cancels a calendar event on a user's behalf, Pieces always asks for confirmation, because a deleted event cannot be restored. Creating, updating or rescheduling events goes ahead without that step. What the assistant can draw on stays under the user's hand: capture can be paused at any time or switched off for individual apps, websites and capture methods, and organization admins choose which features and capture sources members can use at each rollout stage.

Sourcedocs.pieces.app/products/desktop/connectors/google-calendarread 2026-11-09

Organizations assign four roles, Owner, Admin, Write and Read: admins invite members, change roles and manage billing and settings, writers create their own content and readers only view. Admins can connect single sign on through a Descope tenant and auto assign users by company email domain, and accounts support multi factor authentication. Pieces' privacy policy cites least privilege access with periodic reviews, customer managed encryption keys where available on enterprise deployments, a SOC 2 Type II certification and independent penetration testing.

Sourcedocs.pieces.app/products/organizations-and-teams/member-permissionsread 2026-11-09

The Timeline records captured events and every summary Pieces generates in order, with rollups that group a period's activity and filters by event, app and time, and web answers carry citations to their sources. That record follows the user's activity and Pieces' outputs. A step by step log of the assistant's own memory searches, file reads and calendar changes, or of what an MCP client asked for, is not part of it, and the calendar itself is where changes the assistant made show up.

Sourcedocs.pieces.app/products/desktop/timelineread 2026-11-09

The Long Term Memory engine forms memories every two seconds from the focused app, from copied text and, for users who opt in, from meeting audio, and stores them on the user's own device as a searchable timeline. Pieces' own assistant and any connected MCP client read that memory as context in later sessions, alongside a generated persona of the user. Capture can be paused or excluded for apps and websites, and memory can be cleared permanently by time period, capture method or app, or deleted entirely by removing the local data folder.

Sourcedocs.pieces.app/products/core-dependencies/pieces-os/long-term-memoryread 2026-11-09

Memory is captured, processed and stored on the user's own computer, on macOS, Windows or Linux, and nothing leaves it unless the user turns on Personal Cloud or sends a request to a cloud model, which receives only the context scoped to that question. Local models through Ollama keep the whole loop on the device. Enterprise organizations can route cloud requests through their own Anthropic, OpenAI or Gemini accounts instead of the providers Pieces reaches through its routing provider.

Sourcedocs.pieces.app/products/core-dependencies/on-device-storageread 2026-11-09

Eleven ready made summary types come with every plan: Today's Headlines, What's Top of Mind, Morning Brief, Standup Update, Meeting Prep, Day Recap, Time Breakdown, Week Recap, AI Habits, Professional Persona and Collaboration Patterns. Each runs in one click or on a schedule and does its own job, from adding prep blocks to a calendar to breaking a day down by project or mapping who a person works with. Customization templates let users build their own summary types on top of these.

Sourcedocs.pieces.app/products/desktop/single-click-summaries/default-typesread 2026-11-09

Summaries can run without anyone asking. Any built in summary type, such as Morning Brief, Standup Update or Day Recap, can be scheduled daily or on chosen weekdays at a preset or custom time, and the result lands in the Timeline. Schedules fire only while Pieces is open on the computer, and Custom Summary cannot be scheduled. Outside agents reach Pieces on demand through its MCP server from clients such as Cursor, Claude Code and Codex, and the desktop app, CLI and editor plugins take questions directly.

Sourcedocs.pieces.app/products/desktop/single-click-summaries/schedulingread 2026-11-09

Users pick the model for each question, switching between Claude, Gemini, ChatGPT and local models running through Ollama under one plan. Enterprise admins turn on a customized allowlist that sets which providers, and which of each provider's models, members can use, and the choice syncs to every member's Pieces Desktop and PiecesOS. Perplexity handles web search on its own and is not a selectable model.

Sourcedocs.pieces.app/products/organizations-and-teams/settings-modelsread 2026-11-09

Outside agents reach Pieces through its MCP server, which PiecesOS serves on the user's machine at a local endpoint over SSE or streamable HTTP, with one click setup for more than fifteen clients and remote access through mcp-remote or an ngrok tunnel. Its tools query long term memory, search workstream summaries and create new memories, named across the client setup guides rather than in one reference. Pieces offers no public REST API or SDK for its platform; a CLI covers saved materials, chat, cloud models and PiecesOS settings.

Sourcedocs.pieces.app/products/mcpread 2026-11-09

Nothing in Pieces tests or scores the assistant's behavior. There are no test cases, quality checks or comparisons between versions of a summary, and a user who dislikes a result edits the summary's template or asks again. Because the model can be switched per question, a user can rerun the same prompt on another model and compare the answers by eye, and the answers themselves are judged by reading them.

Sourcedocs.pieces.app/products/desktop/single-click-summaries/customization-templatesread 2026-11-09

Pieces reads the screen but never operates it. Vision capture records what is visible in the focused application, and browser history, bookmarks and tabs feed memory, but the assistant does not open pages, click, type or drive a browser or desktop session. Web answers come from Perplexity search results, and calendar changes go through Google Calendar's service rather than an interface the assistant works.

Sourcedocs.pieces.app/products/core-dependencies/pieces-os/long-term-memoryread 2026-11-09