Skip to main content
Glama
supramal1

Cornerstone MCP Server

by supramal1
OverviewSchemaSourceRelated ServersDiscussionsAdmin
Python
Remote

Cornerstone MCP Server

MCP server providing 17+ tools for Claude Code, Claude Desktop, and any MCP-compatible client to interact with Cornerstone memory.

Quick Start

  1. Clone and set up:

    git clone https://github.com/supramal1/cornerstoneMCP.git
cd cornerstoneMCP
bash setup.sh

  2. Edit .env with your API URL and key

  3. Add to Claude Code settings

  4. Start using it:

    • "Remember that the project deadline is March 15" → saves to memory

    • "What do we know about the Google pitch?" → recalls from memory

    • "Forget the old project deadline" → removes from memory

Related MCP server: substrate-memory

Deploy

Deployed to Google Cloud Run (europe-west2). See cornerstone repo for deploy scripts.

Auth

OAuth 2.1 with JWT principal-based access control. Falls back to legacy API key auth.

Configuration

Required env vars: CORNERSTONE_URL, CORNERSTONE_API_KEY, OAUTH_JWT_SECRET, MCP_PUBLIC_URL.

Optional: DISABLE_OAUTH=true skips OAuth advertisement entirely (no /.well-known/oauth-*, no WWW-Authenticate: Bearer ...). Use when this MCP is fronted by a gateway (e.g. obot) that owns user auth and proxies inbound traffic without an Authorization header. Backend calls still authenticate via CORNERSTONE_API_KEY.

Tools

Simple (start here)

Tool

Purpose

Example

remember

Save anything to memory

remember("The budget is $50K")

recall

Search memory

recall("client's email?")

forget

Remove from memory

forget("old_deadline", confirm=True)

Workspace

Tool

Purpose

list_workspaces

See available workspaces

switch_workspace

Change active workspace

get_current_workspace

See which workspace is active

set_default_workspace

Set default workspace

Advanced

Tool

Purpose

get_context

Full context retrieval with metadata

add_fact

Save a structured key-value fact

add_note

Save a freeform note

search

Search across all memory types

list_facts

List recent facts

get_recent_sessions

See recent conversations

list_threads

See conversation threads

report_context_feedback

Rate retrieval quality

Review (MC-2 propose-then-approve)

Tool

Purpose

mc_list_proposals

List agent-proposed facts awaiting human review

mc_promote_facts

Promote proposed facts to current memory

mc_reject_facts

Reject proposed facts (reason required, stored verbatim)

Promote/reject relay the gateway-injected X-Forwarded-User-Email header on the outbound API call so the decision is attributed to the human reviewer. Absent the header (e.g. stdio), the decision is recorded without a forwarded identity.

Most users only need remember and recall. The advanced tools give you more control when you need it.

This server cannot be deployed

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

Maintenance

ActivityInactive
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers