Kept
Archive your AI conversations as local Markdown files with no cloud required, then search, browse, and reuse them via a desktop app and MCP server.
At a Glance
About Kept
Kept is an open-source, local-first tool built by E-Group Labs that captures conversations from popular AI chat providers and writes them to your filesystem as plain Markdown files. It requires no account, no cloud sync, and ships under the MIT license. The public release on GitHub is the full feature set — the project page states there is no premium tier and no gated build.
What It Is
Kept sits between your browser and your local machine. A sideloaded Chromium extension intercepts conversation data from provider API responses (not scraped DOM) and forwards normalized payloads to a desktop app running on localhost. The desktop app — built with Tauri 2, Rust, and TypeScript — writes one Markdown file per conversation with YAML frontmatter into ~/.kept/vault/, indexed in SQLite for full-text search and in CozoDB for a local knowledge graph. The result is a personal archive that any Markdown-aware tool can read: editors, ripgrep, Obsidian, or diff tools.
Supported Providers and Capture Model
Kept currently syncs conversations from ChatGPT, Claude, Gemini, Grok, and Kimi. The extension uses each provider's private API responses rather than scraping rendered pages, which the README notes makes it more stable but still subject to upstream API changes. Capture can be triggered manually with an @kept inline command that opens a palette, or left to run automatically on a configurable interval with SHA-256 deduplication to prevent duplicates.
Local-First Architecture
The entire archive lives on disk under ~/.kept/. Key components include:
- Markdown vault — one
.mdfile per conversation, grouped by provider, diffable and grep-able - SQLite FTS5 index — full-text search across tens of thousands of conversations, filterable by platform, date, and message count
- CozoDB knowledge graph — entities, decisions, and relationships extracted from conversations and linked for graph exploration
- Smart Recall — semantic search using plain-English queries with citations back to source conversations
- BYOK inference — bring your own keys for OpenAI, Anthropic, or OpenRouter, or run local models via Ollama for fully on-device operation
Nothing leaves the machine by default. Cloud API calls only happen when a user explicitly invokes a feature that requires one, such as Smart Recall against a cloud provider.
MCP Server and Integrations
Kept ships kept-vault-server on npm, an MCP server that exposes the vault to any MCP-aware client — Claude Desktop, Cursor, Cline, Claude Code, and others. The server exposes tools for listing, reading, writing, updating, deleting, moving, and grepping vault files. The project page also describes a Claude Code plugin path and a one-line installer for the MCP server. Kept's own agent layer can call any MCP server the user connects, making the vault a two-way integration point in agentic workflows. An Obsidian export command copies the vault into an existing Obsidian vault with reconciliation logic that preserves user edits on re-runs.
Update: v1.0.1
The latest GitHub release is Kept 1.0.1, published on 2026-05-12. The repository was created on 2026-05-02 and last pushed on 2026-05-26, indicating active early development. The Chrome Web Store submission is pending — the extension is currently distributed as a sideloaded unpacked extension. The README notes the Web Store listing will follow once the review queue clears. The project has 90 stars and 5 forks on GitHub as of the last recorded update.
Community Discussions
Be the first to start a conversation about Kept
Share your experience with Kept, ask questions, or help others learn from your insights.
Pricing
Free
Full feature set, MIT licensed, no premium tier or waitlist.
- Capture from ChatGPT, Claude, Gemini, Grok, and Kimi
- Local Markdown vault
- SQLite FTS5 full-text search
- Smart Recall semantic search
- Knowledge graph
Capabilities
Key Features
- Capture AI conversations from ChatGPT, Claude, Gemini, Grok, and Kimi
- Local Markdown vault with YAML frontmatter, one file per conversation
- SQLite FTS5 full-text search with platform, date, and message count filters
- Smart Recall semantic search with plain-English queries and source citations
- Auto-clustering of conversations by topic with no manual tagging
- Knowledge graph of entities, decisions, and relationships via CozoDB
- MCP server (kept-vault-server) for vault read/write/search from any MCP client
- BYOK support for OpenAI, Anthropic, OpenRouter, and local Ollama models
- Obsidian export with reconciliation logic to preserve user edits
- Continue archived conversations in-app against any configured provider
- Claude Code integration for managing skills, memory, and templates
- SHA-256 deduplication for background auto-sync
- PDF and image reading with text extraction into agent context
- Filesystem context: point Kept at any directory to include local files in agent context
- Code execution against the archive with explicit consent prompts
- Weekly digests, project summaries, and decision logs in plain Markdown
- @kept inline trigger for manual capture with palette UI
