FutureClinic agent plugin

Start a FutureClinic chat with a state-licensed US doctor from your AI app, answer FutureClinic's questions, confirm the summary that goes to the doctor, and read your doctor's replies word for word. FutureClinic is cash-pay, for adults, about their own health, and not for emergencies.

One plugin, two formats: it installs in Claude Code and in Codex from this repository, which is also the plugin marketplace. ChatGPT is not supported yet.

FutureClinic's MCP server runs at https://mcp.futureclinic.com/mcp.

What it does

The plugin adds one remote MCP server and one skill. The server's tools act only inside the signed-in patient's own FutureClinic account; no tool takes a patient id.

Tool	Title your AI app shows	What it does
list_chats	List your chats	Read-only. Lists your FutureClinic chats, latest activity first, each with a status line for each thing open in it and a link.
get_chat	Read a chat and your doctor's replies	Reads one chat: where it stands, your doctor's messages word for word, a doctor choice waiting, and any questions or summary waiting on you. It also marks the chat read, as opening it in the app does.
send_message	Write to FutureClinic	Starts a FutureClinic chat with a doctor, or writes in one of your chats, in your own words.
answer_form	Answer your questions	Sends your answers to the questions in your chat.
upload_attachments	Upload attachments	Gives a single-use FutureClinic upload page where you send photos or a video into your chat, as the app's chat takes them, or for a question that asks for them, after you sign in to your own FutureClinic account. The files never pass through your AI app. No PDFs or documents.
cancel_questions	Stop or restart your questions	Only when you ask: cancels your questions, or starts them over where a result offers it. Either drops the answers given so far. Your AI app may ask you to approve it.
restore_questions	Carry on with expired questions	Brings back questions that expired where you left them, answers included.
finish_questions	Send your answers or accept the doctor's plan	Only on your explicit go: sends answers kept unsent on to the doctor, or takes the doctor's plan shown with your questions as the answer. Your AI app asks you to approve it.
confirm_summary	Send your summary to a doctor	Sends the summary you confirmed to a doctor. Your AI app asks you to approve this call.
get_treatment_plan	Read your treatment plan	Read-only. Reads the treatment plan, or response to lab results, your doctor sent in a chat: its words, prescriptions, lab tests or scans, where it stands and what is left to do, with links. It shows the plan only after you have confirmed your home address and ID on FutureClinic. Accepting the plan, and any other change to it, happens in the FutureClinic app or on its website.
get_payment_link	Get the payment link	Shows the doctor choice, FutureClinic's pick and the first five doctors with their profile links, and records yours once you make it, with a link to FutureClinic's payment page when something is left to pay. The call itself pays nothing.
try_again	Send your request to doctors again	Sends a request no doctor took in time, or one you canceled, to doctors again.

The server's tool descriptions, which your AI app reads, are the full rules for each tool; this table is a summary. Each result keeps the ids and handles your AI app passes back (a chat_id, a doctor's profile handle) in a part marked for the AI app only, and every block your AI app shows you word for word ends on its own end line, so no id reaches the text you read. Every tool declares title, readOnlyHint, destructiveHint, idempotentHint and openWorldHint. The server's own instructions, and each tool's description, carry the rules for each call: emergency advice comes first; FutureClinic's safety advice and your doctor's words are passed on unchanged; a new chat starts only on your yes, after your AI app asks once for your whole story (unless you already told it, or in an emergency), which it uses to answer FutureClinic's questions for you to check. The chat skill (plugins/futureclinic/skills/chat/SKILL.md) tells the model how the tools fit together.

Example prompts

• "Can a FutureClinic doctor look at the itchy rash on my arms?"
• "Has my FutureClinic doctor replied yet?"
• "What did my doctor say about the cream, word for word?"
• "What's in my FutureClinic treatment plan?"
• "I want to pay for my FutureClinic request."
• "I'm ready to answer FutureClinic's questions."

What it sends and stores (privacy)

• The plugin contains no executable code: two manifest pairs, one remote MCP server URL and one skill. It stores nothing on your computer.
• Your AI app connects to https://mcp.futureclinic.com/mcp after you sign in to FutureClinic in your browser and agree to share data with that AI app (OAuth 2.0, through FutureClinic's sign-in at account.futureclinic.com). FutureClinic receives only the tool calls your AI app makes. It has no access to your AI app account.
• list_chats status lines carry no health details. get_chat shows your doctor's messages, and get_treatment_plan your doctor's treatment plan. What you send with send_message and answer_form, and what you upload on the page upload_attachments gives, becomes part of your FutureClinic medical record, which FutureClinic keeps for the years state law sets and cannot delete before then.
• You pay only on FutureClinic's payment page in your browser, never in the AI app, by card or, when it covers FutureClinic's pick of doctor, with your FutureClinic credit.
• Whatever you send through these tools also passes through your AI company, under its own terms.
• You can disconnect an AI app at any time in your FutureClinic account settings; FutureClinic then refuses its calls.
• FutureClinic's privacy policy: https://futureclinic.com/privacy-policy. Terms: https://futureclinic.com/terms-of-service.
• For an emergency call 911, or call or text 988 for a mental-health crisis.

Connect (patients)

Patients follow https://futureclinic.com/connect-your-ai once FutureClinic publishes it, which gives one server line per app. Set FutureClinic up once per app: a second setup at the same address, such as this plugin beside a direct line, makes you sign in twice.

Claude (claude.ai, Desktop, mobile) and Claude Code signed in with a Claude subscription: add https://mcp.futureclinic.com/mcp as a custom connector at https://claude.ai/customize/connectors (Add custom connector), or use the "Add to Claude" link on FutureClinic's setup page. It works in Claude Code too, so do not also add the server below.

Claude Code without a Claude subscription:

claude mcp add --transport http --scope user futureclinic https://mcp.futureclinic.com/mcp

Then type /mcp, choose futureclinic and Authenticate, or run claude mcp login futureclinic.

Codex:

codex mcp add futureclinic --url https://mcp.futureclinic.com/mcp
codex mcp login futureclinic

Change no approval settings: FutureClinic's submit step (confirm_summary) asks for your approval on purpose.

Install the plugin from this repository

Use the plugin or a direct line above, never both.

Claude Code:

claude plugin marketplace add futureclinic/agent-plugin
claude plugin install futureclinic@futureclinic

Then type /mcp, choose plugin:futureclinic:futureclinic and Authenticate.

Codex:

codex plugin marketplace add futureclinic/agent-plugin
codex plugin add futureclinic@futureclinic
codex mcp login futureclinic

Troubleshooting

• FutureClinic's sign-in page or your AI app says AI apps aren't available for your account yet. FutureClinic is opening AI apps to patients account by account, and yours does not have them yet. Your chats are still in the FutureClinic app and website.
• Your AI app says the sign-in expired or asks you to authenticate again. In Claude Code type /mcp, choose the FutureClinic server and Authenticate; in Codex run codex mcp login futureclinic; in claude.ai reconnect the connector.
• A tool call says the app's copy of the tools may be out of date. Start a new session (Codex keeps its old tool list until then).
• A payment or upload link expired. Ask your AI app for a new one.
• The upload page asks you to sign in or will not open. It opens only when you are signed in to your own FutureClinic account, on your phone or computer. Sign in as yourself, then open the link again.
• Your summary was not sent because you are in another state. For a request to FutureClinic's doctors: update your address on FutureClinic's website, then confirm again. For questions from your own doctor, when the answer says that doctor isn't licensed in your new state: confirming again will not send them. Update your address, then start a new FutureClinic chat, which finds a doctor licensed there; your chat with that doctor stays locked meanwhile. Your AI app passes on the way FutureClinic gives.
• Anything else. Email help@futureclinic.com.

Layout

.agents/plugins/marketplace.json       marketplace "futureclinic" for Codex (its primary path)
.claude-plugin/marketplace.json        marketplace "futureclinic" for Claude Code (Codex's legacy fallback)
plugins/futureclinic/
  .claude-plugin/plugin.json           Claude Code manifest
  .claude-plugin/icon.png              directory listing icon (the green square logo)
  .mcp.json                            Claude Code server ("type": "http")
  plugin.json                          Agent Plugins 1.0.0 manifest (Codex 0.160+, OpenAI)
  mcp.json                             Agent Plugins server ("type": "streamable-http")
  skills/chat/SKILL.md                 the playbook the model follows
  skills/chat/agents/openai.yaml       OpenAI: the skill's MCP dependency
  evals/                               claude plugin eval cases with MCP mocks
schemas/                               vendored agent-plugins.org 1.0.0 JSON schemas
scripts/validate.mjs                   repository checks (CI)
scripts/validate.test.mjs              one broken copy per check, so none goes quiet (CI)

Codex 0.144 reads .mcp.json, Codex 0.160+ reads mcp.json, Claude Code reads .mcp.json. Keep the two server files identical (same name and URL), and the two marketplace files listing the same plugins at the same paths; scripts/validate.mjs fails when they differ. The plugin name futureclinic is permanent once released.

Develop

pnpm install
pnpm validate                                              # schemas, YAML, directory rules, cross-file checks
pnpm test                                                  # each check still fails when its rule breaks
claude plugin validate . --strict                          # marketplace
claude plugin validate plugins/futureclinic --strict       # plugin
claude --plugin-dir plugins/futureclinic plugin list       # loads in Claude Code

Never use --strict-mcp-config with this plugin: it drops the plugin's server.

pnpm validate also checks the Claude directory's rules that a file can show: a skill frontmatter without allowed-tools (a directory plugin may not pre-answer your permission prompts), no skill text that names the model's memory, past, previous, earlier or prior chats or conversations, chat history or logs, conversation summaries, what it already knows about the user, or the user's files (policy 1.F; a word list, so review still reads the rest), a plugin README of at least 40 words outside code blocks, no system files such as .DS_Store, only remote https:// servers (the plugin runs no local command, so it has no package launcher to pin), and, over the saved tool list, every tool name within 64 characters as Claude Code names it and every tool with title, readOnlyHint and destructiveHint. It fails when a Markdown file in the plugin, this README or docs/ names a tool the saved list does not have: a full mcp__plugin_... name anywhere, or a backticked snake_case name that is no tool, argument or argument value of the saved list. A skill line that lists a tool's arguments, such as - `get_chat` (`chat_id`, optional `wait_s`), may name only that tool's own arguments and values, and calls none optional that the tool requires. It fails, too, when a skill's agents/openai.yaml dependency description names a tool the list does not have, when the skill never names a tool the list has, and when a mock hands out a chat_id that get_chat's mocks have no file for. It checks that hints exist, not that they are right.

Evals

plugins/futureclinic/evals/ holds claude plugin eval cases. They answer the FutureClinic tools from mocks (evals/mocks/futureclinic/), never the real server, and use made-up data only. They call a model on your account, so they run locally before a release, not in CI:

cd plugins/futureclinic
claude plugin eval . --runs 3 --no-publish --max-cost-usd 5

A mock answers every call the same way, except get_chat, which answers from get_chat/<chat_id>.md for the chat asked for, so a case reads the chat its own timeline has; a chat_id with no file there fails the run.

evals/mocks/futureclinic/_tools.json is the server's own tools/list answer, saved as it was sent. When a tool's name, description, schema or annotations change on the FutureClinic server, save the new answer there, then update the skill and the mocks that answer differently, and change the source below. The saved copy, and every mock's wording, is the answer of the FutureClinic server built from its source of 2026-10-23, saved that day, before that source reached production; this line is the one place that records it. Production may still answer differently until its release, which the submission checklist checks before the listing goes in.

Support

Email help@futureclinic.com. Privacy policy: https://futureclinic.com/privacy-policy. Terms: https://futureclinic.com/terms-of-service.

License

MIT, see LICENSE. The FutureClinic name and logo are trademarks of FutureClinic and are not licensed for other use.