Model Context Protocol

Thinkpol MCP

Install Thinkpol in your coding agent, then point the agent back to this page so it uses Thinkpol with the right workflow instead of guessing.

Part 1

For the User

Install Thinkpol in Codex or Claude Code, then explicitly tell the agent to read this page before it starts investigating.

If you have not used these tools before: Codex is OpenAI's coding agent, and Claude Code is Anthropic's coding agent. Choose the one you use below and add Thinkpol there.

Install in Codex

OpenAI's coding agent.

OverviewMCP docs

Add Thinkpol to Codex with the hosted MCP URL:

export THINKPOL_API_KEY=...
codex mcp add thinkpol --url https://api.think-pol.com/mcp --bearer-token-env-var THINKPOL_API_KEY

Install in Claude Code

Anthropic's coding agent.

OverviewMCP docs

Add Thinkpol to Claude Code as a remote HTTP MCP server:

export THINKPOL_API_KEY=...
claude mcp add --transport http thinkpol https://api.think-pol.com/mcp --header "Authorization: Bearer ${THINKPOL_API_KEY}"

What to tell your agent

After installation, explicitly direct your agent to this page. That avoids generic MCP behavior and puts it on the Thinkpol-specific workflow immediately.

Use the Thinkpol MCP server for Reddit intelligence work.
Read https://api.think-pol.com/mcp/docs first.
Then follow the Part 2 agent instructions on that page.
Part 2

For the Agent

Use Thinkpol as a bounded, evidence-oriented investigation tool. Pull direct proof first, then add context only where it sharpens the claim.

Read this page or thinkpol://guides/remote-quickstart first. The hosted Thinkpol MCP server is stateless, so keep track of the ids, timestamps, URLs, and snippets you may need in later turns.

  1. Use typed Thinkpol tools first. Start with search.reddit or subreddit.search, then narrow by time window, search mode, content type, or exact ids.
  2. Use mode=phrase for exact fragments and mode=word for broader topical exploration.
  3. Verify specific accounts with user.get_metadata, user.list_comments, user.list_replies, and user.list_posts.
  4. Add thread context only when needed with content.get_post, content.list_post_comments, content.get_comment, and content.list_comment_replies.
  5. Back every factual claim with returned data: URLs, ids, timestamps, subreddit names, and short snippets. If a conclusion is indirect, label it as an inference.
  6. Keep requests tight. Prefer narrow limits, after or since cursors, and smaller time windows before broadening.
  7. Use search.popularity only for single-term trend counting, not as evidence retrieval. Use thinkpol.get only as a fallback for known allowlisted read-only routes.

Thinkpol also exposes MCP-native guidance through the prompt investigation.plan and the resources thinkpol://guides/remote-quickstart, thinkpol://guides/investigation-workflow, thinkpol://guides/search-semantics, and thinkpol://guides/tool-selection.