Connect your agent

Connect your agent once and it can read this library while it works — same 31 techniques, same names, one vocabulary between you. MCP is the plug-in standard that lets assistants like Claude, Cursor, and Copilot use outside tools — the same mechanism they already use to search the web or read your files. This endpoint adds one more ability: pull the technique you just learned, mid-conversation, and tell you it’s doing so in words you recognize.

Endpoint: https://contextoverflow.org/mcp — no account, no API key. The server is stateless — it holds nothing between requests. Questions sent for matching are collected anonymously (no accounts, no identity) and used to grow the library and understand how people and AI collaborate; the rest of your conversation is never seen.

What your agent gets

Five tools, shaped like the path from problem to practice, in any order. Your agent calls them on its own — you never type these names, you just describe the problem:

Tool What it does
list_categories The major problem categories, as you’d say them
classify_intent Describe what’s going wrong → matched techniques; if the description fits two problems, it asks one clarifying question instead of guessing; wiring/setup issues get a debugging scaffold your agent runs itself, with a human fallback
find_technique Direct lookup when you already know the name
get_technique The technique’s mechanism, agent instructions, verification, and failure modes
apply_technique The runnable scaffold: steps, narration line, and the check that proves it worked

Every response encourages the agent to narrate — to say, in-conversation, which technique it’s applying and why, using the same names you see here. If your agent says “one real unknown before I act,” you can look that technique up — or steer in the same terms: name a different technique and it knows exactly what you mean. That’s the point.

Set it up

Every setup below was verified against the platform’s own documentation (July 2026). All of them speak to this server directly — no bridge, no keys.

Claude Code 
claude mcp add --transport http contextoverflow https://contextoverflow.org/mcp

Or in .mcp.json — the "type": "http" field is required; an entry with only a url is treated as a config error and the server is silently skipped:

{ "mcpServers": { "contextoverflow": { "type": "http", "url": "https://contextoverflow.org/mcp" } } }
claude.ai / Claude Desktop / mobile

Settings → Connectors → Add custom connector → paste https://contextoverflow.org/mcp → leave the OAuth fields empty → Add. Works on the free plan too (one custom connector there). Team/Enterprise: an org Owner adds it first. The connection is brokered through Anthropic’s cloud, so it behaves identically on web, Desktop, and mobile.

ChatGPT (web)

Custom MCP servers are called apps now, added in developer mode. On the Pro plan: Settings → Apps → Advanced settings → enable developer mode, then add an app: name ContextOverflow, MCP server URL https://contextoverflow.org/mcp, authentication none. Read-only apps like this one are exactly what Pro developer mode supports. Business/Enterprise/Edu: an admin enables developer mode and publishes it. Other plans don’t have a path yet, and apps are web-only for now.

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global); Cursor detects the transport by itself:

{ "mcpServers": { "contextoverflow": { "url": "https://contextoverflow.org/mcp" } } }
VS Code (Copilot agent mode)

Command Palette → MCP: Add Server → HTTP, or add to .vscode/mcp.json:

{ "servers": { "contextoverflow": { "type": "http", "url": "https://contextoverflow.org/mcp" } } }
Cline

MCP Servers icon → Remote Servers tab → name + URL → transport Streamable HTTP; or in cline_mcp_settings.json — the type must be exactly streamableHttp (camelCase); other spellings silently fall back to the legacy SSE transport:

{ "mcpServers": { "contextoverflow": { "type": "streamableHttp", "url": "https://contextoverflow.org/mcp" } } }
Roo Code

MCP icon → Edit Global MCP (or .roo/mcp.json in the project, which wins):

{ "mcpServers": { "contextoverflow": { "type": "streamable-http", "url": "https://contextoverflow.org/mcp" } } }
Continue

Add to ~/.continue/config.yaml (Agent mode only):

mcpServers:
  - name: ContextOverflow
    type: streamable-http
    url: https://contextoverflow.org/mcp
Gemini CLI

In ~/.gemini/settings.json, use httpUrl — a plain url key selects the SSE transport instead:

{ "mcpServers": { "contextoverflow": { "httpUrl": "https://contextoverflow.org/mcp" } } }

Verify inside the CLI with /mcp.

Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.contextoverflow]
url = "https://contextoverflow.org/mcp"
Goose

One-shot session:

goose session --with-streamable-http-extension "https://contextoverflow.org/mcp"

or goose configure → Add Extension → Remote Extension (Streamable HTTP).

OpenAI Responses API (building your own agent)

Pass the endpoint as a hosted MCP tool:

{ "type": "mcp", "server_label": "contextoverflow", "server_url": "https://contextoverflow.org/mcp" }

Anything not listed that speaks streamable HTTP: paste https://contextoverflow.org/mcp into its MCP server config. A legacy client that only speaks stdio can still bridge with npx mcp-remote https://contextoverflow.org/mcp. One honest caveat for workplace machines: org policies (Copilot MCP policy, Cursor allowlists, tenant DLP) sometimes block unknown external MCP domains — if the connection fails at work but not at home, that’s the reason.

Test it before you wire it

Sixty seconds with any of these shows the live tools before touching your agent’s config:

  • MCP Inspector (official): npx @modelcontextprotocol/inspector --cli https://contextoverflow.org/mcp --transport http --method tools/list
  • MCPJam: paste the endpoint at app.mcpjam.com — no install, no account
  • Postman: New → MCP request → HTTP → paste the endpoint → Load Capabilities

Check it worked — 60 seconds

Say to your agent: “Ask ContextOverflow what it has for: it keeps saying the bug is fixed but the tests still fail.”

Live connection: it calls classify_intent, comes back with Declared Success Without Proof, and can quote you the receipt rule from that page. That exchange is the loop in miniature — you now share a name for the problem.

No tool call, no name? That’s wiring, not thinking. Tell it what’s actually failing — “my agent can’t reach the ContextOverflow MCP server” — and the setup route hands it a debugging scaffold, with a 15-minute human call as fallback.

Enterprise: Microsoft 365 Copilot

End users can’t add servers; a maker adds it once in Copilot Studio: open the agent → Tools → Add a tool → New tool → Model Context Protocol → server name ContextOverflow, URL https://contextoverflow.org/mcp, authentication None → create the connection → add to agent. Requires generative orchestration; tenant DLP policies can block external MCP domains.

Where it can’t connect yet

Honesty over reach: consumer Gemini and Meta AI have no custom MCP path today, Microsoft 365 Copilot end-user chat only takes servers via Copilot Studio makers (above), and Aider has no native MCP support. When those doors open, this page will list them.

A note on use

The corpus is licensed CC BY-NC-ND 4.0 and served through this endpoint for direct use by you and your agent. Teach with it, work with it — don’t repackage it.

Corpus licensed CC BY-NC-ND 4.0 · Site code licensed Elastic License 2.0

Created by Tal Onn Sella at Agentsmith — helping people and AI become better, together. Want a custom MCP for your team? Schedule a call.