AgentTune MCP & portable AI preferences across every agent.
By Bernard Huang · Updated
Add https://agent-tune.com/mcp as a remote Streamable HTTP MCP server in Claude Code ('claude mcp add --transport http agenttune https://agent-tune.com/mcp'), Claude Connectors, Cursor, or any MCP client—no API key or account required. Your agent gets six read-only tools (list_tunings, get_tuning, get_test_spec, list_resources, get_resource, and get_free_tools) to fetch all 43 MIT-licensed tunings, administer open questionnaires locally, and merge bounded preference blocks into CLAUDE.md or AGENTS.md safely.
On this page
One-command setup across Claude Code, Claude Web, Cursor, and Codex.
The AgentTune MCP endpoint at https://agent-tune.com/mcp is stateless, read-only, CORS-enabled for trusted AI origins, and requires zero API keys or login (protocol versions 2025-06-18, 2025-03-26, and 2024-11-05).
| Client / Agent | Exact connection command or config |
|---|---|
| Claude Code CLI | claude mcp add --transport http agenttune https://agent-tune.com/mcp |
| Claude Web & Desktop Connectors | Settings → Connectors → Add custom connector → URL: https://agent-tune.com/mcp |
Cursor (.cursor/mcp.json) | {"mcpServers":{"agenttune":{"url":"https://agent-tune.com/mcp"}}} |
| OpenAI Codex CLI / MCP clients | Point Streamable HTTP client at https://agent-tune.com/mcp, or tell the agent to read https://agent-tune.com/SKILL.md. |
| Any web-fetching agent (no MCP needed) | Read /SKILL.md, /llms.txt, or fetch any raw tuning at /library/{system}/{slug}.md. |
What the 6 MCP tools and machine endpoints expose.
| MCP tool | Inputs | What it returns |
|---|---|---|
list_tunings | Optional system (mbti, enneagram, disc, attachment, ocean) | All 43 MIT-licensed profiles with slugs, codes, one-line blurbs, SHA-256 revisions, and raw Markdown URLs. |
get_tuning | system + slug (e.g. mbti / intj, disc / d-dominance) | Clean Markdown communication rules in structuredContent.body separated cleanly from install metadata. |
get_test_spec | test (mbti, enneagram, disc, attachment, or big-five) | Versioned questionnaire items, anchors, and deterministic local scoring rules—so answers stay 100% local and never cross the network. |
list_resources | Optional kind (guide or research) and query | Searchable index of all setup guides and empirical AI personality studies. |
get_resource | id (e.g. claude-personality, ai-personality-five-models-2026) | Full Markdown mirror and provenance metadata for any guide or research article. |
get_free_tools | None (zero personal input accepted) | Catalog of local browser utilities and the window.AgentTuneTools JavaScript text-processing API. |
Machine-readable discovery cards live alongside the MCP endpoint at /.well-known/mcp/server-card.json, /SKILL.md, /llms.txt, and /library/index.json.
Safe, non-destructive installation with bounded markers.
Handing an autonomous coding agent a raw system prompt and saying "update my CLAUDE.md" risks overwriting build commands or repo rules. AgentTune enforces two strict safety guardrails across every tuning file and tool:
- Explicit user confirmation: Every tuning file carries
does not authorize installationin its contract—fetching a file never grants write permission to~/.claude/CLAUDE.md,~/.codex/AGENTS.md, orSoul.mdunless you explicitly ask the agent to write it and approve the diff. - Bounded HTML markers: Preferences sit inside
<!-- agenttune:preferences:start -->and<!-- agenttune:preferences:end -->markers (the exact lossless merge contract tested in our Instruction Converter). Updating or removing your communication style never touches a single byte of surrounding project setup.
Three prompts to run right after connecting AgentTune MCP.
Or ask your agent to administer the fastest 16-item assessment inline while keeping your answers local:
Questions people ask.
Does the AgentTune MCP server require an API key, signup, or payment?
No. https://agent-tune.com/mcp is completely free, stateless, and requires zero API keys or authentication.
Does AgentTune MCP store my conversations, code, or questionnaire answers?
No. By design, the MCP endpoint is strictly read-only and never accepts personal text or questionnaire answers over the network: get_test_spec returns the public scoring specification so your agent scores answers locally in your session.
What if my AI tool doesn't support remote MCP servers yet?
Point any web-capable agent at https://agent-tune.com/SKILL.md or https://agent-tune.com/llms.txt, or copy a bounded Markdown block manually from agent-tune.com/tools/instruction-converter.
Will applying an AgentTune profile overwrite my existing CLAUDE.md or AGENTS.md?
No—when merged inside <!-- agenttune:preferences:start --> and <!-- agenttune:preferences:end --> markers, only the communication block between the markers is added or updated, preserving every existing build, lint, and architecture instruction byte-for-byte.
What changed.
- Published covering remote Streamable HTTP MCP setup (v1.3.0), SKILL.md discovery, and bounded marker merges.