What it is
kaRkas is a project memory for AI tools. You map your product β jobs-to-be-done, milestones, models, stack, brand, design β and your AI reads it through the Model Context Protocol.
No copy-paste. No re-explaining the codebase. Every Cursor chat, every Claude Code session starts already briefed.
Install
Run one of these in your project root:
# Cursor
curl -sSL https://app.karkas.io/connect/cursor?token=YOUR_TOKEN | sh
# Claude Code
curl -sSL https://app.karkas.io/connect/claude-code?token=YOUR_TOKEN | sh
# Cline, Continue, or any other MCP client
curl -sSL https://app.karkas.io/connect/generic?token=YOUR_TOKEN | shGrab the token in Settings β MCP, or hit Create access tokenon any project's Build tab.
Scripts are server-signed and tokens are hashed in the database. Lose one? Revoke it and make a new one.
Tools
Loading tool catalogβ¦
Resources
Server-streamed reads for tools that prefer resources over tool calls.
karkas://briefingβ your project, distilled into one markdown document. Mission, models, stack, brand, design, current milestone, next task, recent decisions.karkas://project/contextβ raw bundle for tools that want one payload.karkas://projectsβ index of projects this token can see.
Skills
Skills are prompt templates. The install script drops yours into the agent's native skill folder so they load automatically:
- Cursor β
.cursor/rules/skill-<slug>.mdc - Claude Code β
.claude/skills/<slug>.md
Agents can also fetch skills at runtime with list_skills and get_skill.
Scopes
Every tool requires a scope. New tokens get the safe defaults: every read, plus write:scaffold, write:notes, write:tasks, write:milestones, write:quality.
Opt-in only
read:user_memoryβ your global AI memory.write:user_memoryβ append a short preference.read:other_chatsβ read sibling chat threads in the project.
Toggle these in Settings β MCP β Default scopes. New tokens inherit your selection.
Manual setup
If you'd rather wire the config yourself:
Cursor β .cursor/mcp.json
{
"mcpServers": {
"karkas": { "url": "https://app.karkas.io/mcp?token=YOUR_TOKEN" }
}
}Claude Code β ~/.claude/settings.json
{
"mcpServers": {
"karkas": {
"type": "http",
"url": "https://app.karkas.io/mcp?token=YOUR_TOKEN"
}
}
}