@blanklogic/refinery-mcp
v0.2.5
Published
Clean context before AI sees it — BlankLogic Refinery MCP for Cursor and Claude Code (local, no upload).
Maintainers
Readme
@blanklogic/refinery-mcp
MCP server that de-bloats code and redacts secrets for Cursor, Claude Code, and other MCP agents.
Agents get cleaner context and fewer leaked secrets — without sending your repo to BlankLogic servers. Runs locally over stdio.
What's new in 0.2.5
- Coordinated release with Refinery browser app and BYOK Proxy — same
@blanklogic/[email protected]engine. - README and site copy focused on MCP-first install (copy-paste config below).
- Expanded release test gate (CLI, MCP full suite, community stats).
- 512 KB max input · tools:
refine_code,sanitize_secrets,refine_full— see TOOL_CONTRACT.md. - Optional telemetry:
BLANKLOGIC_MCP_TELEMETRY=0to disable.
Saved tokens or caught a secret?
If BlankLogic trimmed noise, redacted secrets, or cut token waste for you, a quick review helps others find it:
- Cursor / Open VSX: Rate BlankLogic Refinery on Open VSX
- VS Code: Leave a review on the Marketplace
- Feedback: blanklogic.io/#support
Install
npx -y @blanklogic/refinery-mcpCursor / Claude Code — paste this
Copy into Cursor Settings → MCP (or Claude Code MCP config):
{
"mcpServers": {
"blanklogic-refinery": {
"command": "npx",
"args": ["-y", "@blanklogic/refinery-mcp"]
}
}
}Restart the agent session, then ask: “Use refine_full on this pasted code before sending it to the model.”
Tools
| Tool | Description |
|------|-------------|
| refine_code | De-bloat input; return cleaned output + stats |
| sanitize_secrets | Redact obvious secrets; return counts |
| refine_full | De-bloat then sanitize (recommended) |
Example response:
{
"tool": "refine_full",
"output": "API_KEY=[REDACTED]",
"stats": {
"bytesIn": 128,
"bytesOut": 18,
"tokensSaved": 12,
"secretsRedacted": 1
}
}Who this is for
- Cursor and Claude Code users who want agents to clean payloads automatically
- MCP directory listings (Smithery, mcp.so, PulseMCP, etc.)
- Teams standardizing context hygiene across agent workflows
Related BlankLogic tools
| Product | Link |
|---------|------|
| CLI | @blanklogic/refinery |
| VS Code / Cursor extension | VS Marketplace · Open VSX |
| AI cost proxy | @blanklogic/proxy |
| Web app | blanklogic.io/refinery |
Local development (monorepo)
npm install
npm --workspace @blanklogic/refinery-mcp run smokeDev config with absolute path:
{
"mcpServers": {
"blanklogic-refinery": {
"command": "node",
"args": ["/path/to/blanklogic-site/packages/mcp-server/src/index.js"]
}
}
}Tool contract
See TOOL_CONTRACT.md for canonical tool names, the input argument, and the 512 KB limit.
Cloud credits (optional)
MCP tools run locally over stdio. For server-side cleanup (larger payloads, Cloud Refine, or Cloud AI), use the Refinery web app:
blanklogic.io/refinery/#cloud-credits — Beta Token Pack ($5 / 500 runs) or Founder pass ($50).
Privacy
Runs locally over stdio. Does not call Stripe, Ollama, or third-party LLM APIs for MCP tool calls. Cloud credits apply only to the web flow linked above.
By default, sends anonymous aggregate usage (bytes/timing only, never your input) to BlankLogic public stats. Disable with BLANKLOGIC_MCP_TELEMETRY=0. See TOOL_CONTRACT.md.
Support
- Homepage: blanklogic.io
- Web app: blanklogic.io/refinery
- Bug reports: blanklogic.io/#bugs
- Directory tags:
mcp,cursor,claude-code,ai-agents,context-hygiene,secret-redaction
