Silt
In developmentSilt keeps LLM-maintained project wikis aligned with source documents, meeting transcripts, and notes. I have used personal versions of Andrej Karpathy’s LLM-maintained wiki approach across research and work. Silt makes the LLM-maintained wiki method practical with continuously changing source material.
The problem
In my day job, I move between projects often. Project material includes Google Docs such as statements of work, Jira tickets, Read.ai meeting transcripts, and many other sources. I use an LLM-maintained wiki to manage and query project context, but source material continually changes. Documents are edited, tickets change, and new transcripts arrive.
Keeping project material current limited effective use of the wiki method. Re-reading every source or manually explaining document changes creates an unsustainable workflow.
Silt manages source material for the wiki method. A vault contains raw/ source material, wiki/ pages written by an agent, and a CLAUDE.md with agent instructions. Silt writes only to raw/ and never edits wiki pages. The application keeps source material current and identifies required wiki updates.
Agent workflow
Wiki-writing agents connect to Silt’s local HTTP server through MCP. During each work session, an agent can check for pending changes, request a manifest of new or modified files, read project files from the vault, update the wiki, and acknowledge the incorporated change set.
An ingestion watermark replaces a generic clearance action. Every payload includes an ingestion mark for a defined file set. An acknowledgement cannot discard files that arrive during wiki updates. The ingestion workflow makes wiki updates repeatable agent tasks rather than manual handoffs.
Architecture
Silt separates the ingestion engine from the desktop client. The Go engine runs a local HTTP API and MCP server, coordinates connectors and transformations, persists project state in SQLite, and stores source material in the vault. The native macOS desktop application uses Rust and Dioxus to manage projects, integrations, sources, updates, and change visibility.
Silt supports local folders, Read.ai transcripts, and Google Docs. Connectors use source-appropriate change detection: Google Drive versions for individual documents and source metadata for other material. The Go engine uses a native DOCX-to-Markdown transformer, preserves raw files in the vault, and tracks artifacts for incremental payloads.
Source-specific transformer configuration defines an external executable path and arguments for project-specific transformations, such as Python scripts. The extension point avoids a fixed built-in transformation catalog and enables specialized source workflows.
OAuth integrations run through the desktop application. The local agent-facing MCP server provides four operations: status checks, pending payload retrieval, ingestion acknowledgements, and source updates.
Stack
Go powers the ingestion engine, HTTP API, connectors, and MCP server. SQLite stores project configuration, source cursors, artifact metadata, run history, and ingestion watermarks. Rust and Dioxus power the native macOS desktop client. Silt uses MCP to connect local wiki agents to source-update workflows.