Pieces for Developers

Runs a background daemon (PiecesOS) that monitors OS-level events across all applications in real-time, capturing code snippets, browser tabs, chat messages, documents, and highlights without user intervention. The Workstream Pattern Engine ingests millions of micro-events and routes them through on-device classification models (TF-IDF, SVMs, LSTMs, RNNs) to automatically detect code, extract metadata, flag sensitive data (PII/credentials), and associate context (source app, timestamp, related files/tabs). Captured data is stored locally in a proprietary database with optional cloud sync via Pieces Drive.

Indexes all captured snippets, documents, and activity with vector embeddings, enabling semantic search via natural language queries. Users can search across 9 months of personal context and filter by time-based queries (e.g., 'code I wrote last Tuesday', 'snippets from the past week'). The search engine ranks results by relevance and associates results with the 'bigger picture' — implied relationship graph linking snippets to related chats, tabs, and documents. Queries are processed locally by default; optional cloud search available via Pieces Drive.

All captured context and snippets are stored in a local, proprietary database on the user's machine by default. Cloud sync via Pieces Drive is optional and user-controlled — users can enable/disable sync at any time. No data is transmitted to cloud unless explicitly enabled. Local storage uses vector embeddings for semantic search and supports 9 months of retention with automatic deletion of older data.

Uses hardware acceleration (GPU, NPU, or CPU optimization — specific method undocumented) to run on-device ML models (TF-IDF, SVM, LSTM, RNN) in real-time as context is captured. Models classify code, detect language, associate context, and flag sensitive data without cloud transmission. Hardware acceleration enables low-latency inference on millions of micro-events per day.

On-device models automatically detect programming language, framework, and code type (function, class, snippet, etc.) from captured code. Extracted metadata is stored with the snippet and used for search, filtering, and context association. Detection runs in real-time without user input or cloud transmission.

Optional cloud sync service (Pieces Drive) that synchronizes local memory to cloud storage for backup, multi-device access, and team collaboration. Users can enable/disable sync at any time. Sync mechanism (incremental, full, real-time) is undocumented. Team collaboration via Pieces Drive enables shared memory across team members with role-based access control.

Provides an AI copilot that accepts user queries and automatically injects personal context (saved snippets, activity history, related documents) before routing to a user-selected LLM backend. Supports Claude (4 Sonnet, Opus), Gemini 2.5, OpenAI models, and Ollama (local). The copilot 'knows what you know, not just what the LLM knows' — meaning it personalizes responses based on your saved code, patterns, and project context. Integrates via MCP (Model Context Protocol) server built into PiecesOS, enabling direct injection into Claude, GitHub Copilot, Cursor, and Goose.

Accepts saved code snippets and applies transformations: change programming language, improve readability, optimize performance, or refactor for specific patterns. Transformations are executed by the selected LLM with personal context injected, enabling suggestions that align with your coding style and project patterns. Output can be previewed, edited, and re-saved to memory.

Generates custom shareable links for individual code snippets, enabling developers to share saved code with teammates or in documentation without exposing the full Pieces memory. Supports export to GitHub Gists for version control integration and public sharing. Links are generated on-demand and can be revoked. Export to Gist includes snippet metadata (language, description, tags).

Automatically organizes captured activity into workflow summaries — high-level overviews of work done during a session or time period. Summaries are generated from the captured context (code, chats, tabs, documents) and indexed for smart search. Users can review what they worked on, when, and with what context. Summaries are stored in memory and linked to related snippets.

On-device classification models (TF-IDF, SVM, LSTM) automatically detect sensitive data patterns (API keys, credentials, PII, passwords) in captured code and flag them for user review. Flagged items are marked in the UI but not automatically redacted or encrypted. Users must manually review and delete flagged items. Detection runs in real-time as code is captured, without cloud transmission.

Enables teams to share a collective memory of code snippets, documents, and activity across team members. Team members can search and access shared context, with role-based access control (mechanism unclear — likely admin/member roles). Shared memory is synced across team members via Pieces Drive (cloud sync). Teams plan includes priority support and custom LLM integration (bring-your-own or OpenAI/Anthropic/Ollama).

Provides plugins for VS Code and Chrome (and other IDEs/browsers, specific list unclear) enabling one-click saving of code snippets directly from the editor or browser. Plugins integrate with the desktop app via local IPC, capturing code, file context, and browser tab context automatically. Plugins also enable quick access to saved snippets and copilot assistance from within the IDE/browser.

Pieces includes a built-in MCP server that exposes personal context (saved snippets, activity history) to MCP-compatible tools: Claude, GitHub Copilot, Cursor, and Goose. The MCP server runs on the local machine and injects context into copilot queries without cloud transmission. Enables seamless integration of personal context into existing copilot workflows without switching tools.