Krusch Context MCP
A local MCP server with five verbs that keeps project decisions across sessions: store, retrieve, supersede and audit diffs.
- 83 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The server writes a database and an `.env` file into the project directory and reads the diff before a commit
- Saved decisions and rules remain in the local database and can hold sensitive project details
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/krusch-context-mcpDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The server writes a database and an `.env` file into the project directory and reads the diff before a commit
- Saved decisions and rules remain in the local database and can hold sensitive project details
This is third-party code. Review the repository files before installing.
What it does
Krusch Context MCP gives an agent project working memory and keeps it in a local SQLite database `.agent/context.db`. The agent uses five verbs: `krusch_context_retrieve` pulls active memories and rules within a token budget with citations, `krusch_context_remember` stores facts and rules and catches near-duplicates (similarity from 0.85), `krusch_context_revise` supersedes stale knowledge with lineage or invalidates a rule with a mandatory reason, `krusch_context_nudge` checks a proposed diff against active invariants and `krusch_context_health` shows counts and storage state. The recommended order: call `retrieve` at session start, record decisions through `remember` during work and run `nudge` before a commit, which returns at most three findings. `npx krusch-context-mcp init` prepares the database and `.env` and prints settings for Cursor, Claude Code and Claude Desktop.
Who it is for. Developers working with an agent in Cursor or Claude Code who lose project agreements with every new session.
Good fit when
- An agent in a new session should immediately know accepted architecture decisions and active constraints
- An accepted decision became stale and must be replaced while keeping the history rather than erased
- Before a commit you want to check the diff against project rules
- The diagnosis of a found bug and its lesson should be saved so the agent does not retread the path
Not a fit when
- You need code symbol search, call graphs or Git history traversal: other tools such as the IDE or `krusch-git` do that
- You need a cloud memory service for a team: the server runs locally on your machine
- You need legal checks, document ingestion or contract graphs: the author moved them to separate servers
Example request
Call retrieve with a wildcard and tell me which decisions and constraints are active in this project, then record the API error format decision and check the diff with nudge.MCP
- Transport
- stdio
- Authentication
- not required
Limitations
Node.js 22 or newer is required: the database is created through the built-in `node:sqlite`. Storage is local and sits in the project, so memories do not move between machines without your own sync. `nudge` runs before a commit or manually and rejects an every-turn mode. The many benchmark figures in the README, including DolphinBench, LoCoMo and BEAM, are published by the author with audit logs and one submission is still queued, so check them against primary sources. Stale memories are reviewed on a 30-day TTL.
How to disable. Remove the `krusch-context` entry from the Cursor or Claude Desktop config or run `claude mcp remove krusch-context`. Delete the `.agent/context.db` memory database and the created `.env` in the project separately if you no longer need the memory.
FAQ
How do I connect the server to Claude Code?
Run `claude mcp add krusch-context -- npx -y krusch-context-mcp`. For Cursor add a `krusch-context` server with the `npx` command and the arguments `-y krusch-context-mcp` to `.cursor/mcp.json`.
What does the init command do?
It checks Node 22, creates `.agent/context.db`, clears stale MCP tool schemas from IDE caches, prepares `.env`, seeds initial context, runs a health check and prints ready settings for Cursor, Claude Code and Claude Desktop.
How does the server handle duplicates?
`remember` catches near-duplicates at similarity from 0.85 and proposes `revise` with the `supersede` action. It does not block contrasting rules.
What does nudge check?
A proposed diff against active project invariants. It runs before a commit or manually by default and returns one to three findings.
Is an external service or Docker needed?
No. Storage is SQLite by default and everything runs locally with no cloud dependencies.
Discussion
No questions or reviews yet.
Related
MCP serversAn MCP server that keeps large tool output out of the agent's context window and stores session memory in SQLite
An MCP server that preserves Claude Code context between sessions: decisions, checkpoints and a file cache kept in local SQLite.
Local memory for agents: decisions from browser chats reach Claude Code, Cursor and Windsurf through MCP
kaeru is an MCP server with an agent memory graph: ideas, outcomes and sources across sessions, local, with a team cloud option.
Local memory for agents: stores conversations and project files verbatim and retrieves them with semantic search over MCP
A CLI and MCP integration keeping agent decisions and checkpoints in local SQLite so work survives a session change.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Krusch Context MCP (this card) | no | not required | 83 | |
| ICM | no | 621 | a local CLI stores project decisions and conventions in SQLite and returns relevant entries without checking diffs | |
| CORE: memory and tasks for Claude Code | no | 2,019 | a Claude Code plugin with a knowledge graph and tasks, heavier and tied to one agent |
README in short
The page opens with a banner and a summary of memory benchmark results, then explains what the server is not and describes a one-command quick start. A table of five verbs, a session-lifecycle usage diagram, setup for Cursor, Claude Code and Claude Desktop and detailed sections on external benchmarks with a verification taxonomy follow.