Using PaPut MCP with Coding Agents

This page introduces the steps for saving knowledge gained while coding in Claude Code or Codex to PaPut.

PaPut MCP tools are API-backed. When you want to review local Claude / Codex session files, the AI client reads those files directly through the paput-harvest skill and then submits extracted candidates through PaPut MCP.

The easiest way to set up Claude Code / Codex is the plugin. This page covers the manual setup (MCP config + setup-ai).

Prerequisites

• Node.js / npx is available
• You have a PaPut account
• You can use Claude Code or Codex

OAuth login

Remote HTTP Mode uses the MCP client OAuth flow. Configure the Remote HTTP URL and sign in from the client when prompted.

MCP configuration

Use the fixed Remote HTTP URL below. To pin tool calls to a skill sheet project, set its alias through the X-PaPut-Project-Alias header instead of a URL query.

Claude Code

Add this JSON configuration and replace <alias>.

"paput": {
  "type": "http",
  "url": "https://mcp.paput.io/mcp",
  "headers": {
    "X-PaPut-Project-Alias": "<alias>"
  }
}

Codex

Add this to the project's .codex/config.toml and replace <alias>.

[mcp_servers.paput]
url = "https://mcp.paput.io/mcp"
http_headers = { "X-PaPut-Project-Alias" = "<alias>" }

Codex can also use env_http_headers when the alias must come from an environment variable. Clients that cannot set custom headers can connect to PaPut, but cannot pin the connection to a project.

setup-ai

Add PaPut skills and global usage rules for Claude / Codex.

npx -y paput-mcp setup-ai

This command adds PaPut skills for harvesting knowledge from past local sessions reviewed by the AI client, extracting knowledge candidates, and confirming before saving. It also adds usage rules that detect reusable technical knowledge at the end of work and add it to pending candidates.

To update existing settings, add --force.

npx -y paput-mcp setup-ai --force

To target Claude only, add --claude-only.

npx -y paput-mcp setup-ai --claude-only

Automatic pending candidate flow

After setup-ai, PaPut knowledge capture rules are added to the global instructions used by Claude Code or Codex. When work is completed, a problem is solved, or a design decision is settled, the AI checks for reusable technical knowledge and adds non-duplicate, non-sensitive, non-project-specific candidates to pending.

Normally, you do not need to run /paput-capture manually every time. Use /paput-capture only when you explicitly want to extract candidates from a specific conversation or topic.

Example skills you can use

• /paput-harvest: In a local-file-capable AI client, inspect unprocessed local sessions and prepare reusable knowledge candidates
• /paput-capture: Manually add knowledge candidates from the current conversation or a specified topic to pending
• /paput-save: Review pending candidates and save only the ones you explicitly approve to PaPut

Recommended workflow

/paput-harvest
Do your work
Run /paput-save when needed

• Useful knowledge is usually added to pending automatically while you work
• Before saving, use /paput-save to review candidates and filter out duplicates or project-specific information
• Use /paput-capture only when you want to explicitly capture a specific topic

Notes

• PaPut MCP itself does not read local files
• Past-session import depends on the AI client's own local file access
• No API key is needed; OAuth is used
• Generated skills are saved under ~/.paput
• If you change the MCP configuration on the Claude Code or Codex side, a restart or reload may be required