> ## Documentation Index > Fetch the complete documentation index at: https://docs.profy.cn/llms.txt > Use this file to discover all available pages before exploring further. # MCP Tools > The 22 tools exposed by the remote MCP server: parameters, semantics, and known defects # MCP Tools Profy runs a remote MCP (Model Context Protocol) server so you can go from distillation to configuration to publishing to monetisation inside Codex, Cursor, Claude Code, or any MCP-capable IDE — without leaving the editor. ``` https://mcp.profy.cn/mcp ``` Transport is Streamable HTTP and **stateless** (no session management). ## Setup ```json .mcp.json theme={null} { "mcpServers": { "profy": { "type": "streamable-http", "url": "https://mcp.profy.cn/mcp" } } } ``` Authentication reuses the platform login session in two forms: * `Authorization: Bearer ` * Browser cookie (for embedded-browser IDE flows) Calling an identity-bound tool while unauthenticated errors out. It does **not** silently degrade to anonymous access. ## Full tool table 22 tools, all thin proxies over existing platform services with no independent business logic. ### Identity (2) | Tool | Parameters | Purpose | | --------------------- | -------------------------------------------- | ------------------------------------ | | `get_creator_profile` | none | Read your creator status and details | | `apply_creator` | `displayName`, `bio`, `accountType`, `email` | Apply to become a creator | `apply_creator` enters human review. The outcome arrives by in-app notification; there is nothing to poll on the MCP side. ### Agent configuration (4) | Tool | Parameters | Purpose | | ------------------- | ----------------------------------------------------- | --------------------------- | | `create_expert` | `name`, `description`, `category` | Create an agent draft | | `update_expert` | `identifier` plus `name` / `description` / `category` | Update configuration fields | | `get_expert_detail` | `identifier` | Get one agent's detail | | `list_my_experts` | none | List all agents you created | `update_expert` covers basic metadata only. **Prompt layers (GUARD / persona / soul / agent), subagents, and case demos are not on the MCP surface** — they are either intellectual property or structurally too intricate to edit blind from an IDE. Those go through Studio. ### Skills (3) | Tool | Parameters | Purpose | | ------------------------- | ------------------------- | ----------------------------------------- | | `install_skill` | `identifier`, `skillName` | Install a marketplace skill onto an agent | | `list_expert_skills` | `identifier` | List installed skills | | `list_marketplace_skills` | `keyword` | Browse the skill marketplace | **`install_skill` is a stub and installs nothing.** The current implementation writes a log line and returns `{"success": true, "installed": ""}`; the installation service is not wired. The call succeeds, the response says success, and `list_expert_skills` afterwards shows no change. This is the most dangerous defect shape — a failure wearing a success mask. You get no error, only confusion about why the skill has no effect. Install skills through Studio or `POST /api/expert-skills/install`. ### Billing and publishing (3) | Tool | Parameters | Purpose | | ----------------------- | ---------------------------------------------------------------------- | --------------------- | | `set_billing` | `identifier`, `billingType`, `price`, `freeTrialPerUser`, `allowTrial` | Set pricing and trial | | `publish_expert` | `identifier` | Submit for review | | `get_submission_status` | `identifier` | Check review status | `set_billing` takes `billingType` as `METERED | ONE_TIME`, but neither name **maps to the real billing model**: * `METERED` → writes price 0 → the agent becomes **free** * `ONE_TIME` → writes `price` → the agent becomes **unlock-on-purchase** The platform has only `FREE` and `UNLOCK`; there is no metered agent. So passing `METERED` expecting usage-based charging yields a fully free agent. This is naming drift between the MCP surface and the underlying model; the underlying model is authoritative. `freeTrialPerUser` defaults to 3 and `allowTrial` defaults to `true`. Full price and trial constraints are in [Pricing and Billing](/en/creators/pricing-and-billing). `publish_expert` **submits for review** rather than going live; a published agent stays online while its new version queues. ### Distillation (2) | Tool | Parameters | Purpose | | ----------------------- | ----------------------------------------------- | ----------------------------------------------------- | | `start_distillation` | `topic`, `materials[]` (`filename` + `content`) | Start the distillation interview with local materials | | `continue_distillation` | `sessionId`, `message` | Answer the AI's follow-up and advance | Distillation is multi-turn: `start_distillation` returns a `sessionId`, and every subsequent answer goes through `continue_distillation` carrying that id. Both tools call the invoke endpoint through an SSE collector internally, folding a stream into one synchronous return — so a single call can take a while. `materials` carries file contents directly (not paths), so local files in your IDE feed distillation without a separate upload. ### Testing (2) | Tool | Parameters | Purpose | | ------------------ | ----------------------- | ---------------------------------------- | | `test_expert` | `identifier`, `message` | Send one message and read the reply | | `open_expert_chat` | `identifier` | Get a browser URL to chat with the agent | `test_expert` really executes and **really bills**. For a free look, use `open_expert_chat` and try it in the browser. ### Navigation (1) | Tool | Parameters | Purpose | | ----------------- | ------------------------------------------------------- | ------------------------------------------- | | `open_profy_page` | `page` (`studio` / `chat` / `marketplace` / `earnings`) | Get the URL of one of four checkpoint pages | It returns a URL only; whether to open a browser is the IDE's decision. ### Self-evolution (3) | Tool | Parameters | Purpose | | ---------------------------- | -------------- | ------------------------------------ | | `list_evolution_suggestions` | `identifier` | List pending improvement suggestions | | `accept_evolution` | `suggestionId` | Accept one suggestion | | `get_evolution_stats` | `identifier` | Evolution statistics and trends | `accept_evolution` **takes effect immediately with no re-review** — skill documents are read at runtime. That is also why you should read the suggestion before accepting. The mechanism is in [Self-Evolution](/en/creators/darwin-evolution). ### Earnings (2) | Tool | Parameters | Purpose | | -------------------- | ----------------------------------------- | ---------------------------------------------------------- | | `get_earnings` | none | Earnings summary (diamond balance and pending withdrawals) | | `browse_marketplace` | `keyword`, `category`, `page`, `pageSize` | Browse published agents | `browse_marketplace` is the only tool in the table that does not require an identity. ## A typical flow Building an agent from scratch inside your IDE: ``` 1. get_creator_profile confirm creator status 2. apply_creator (if not onboarded) 3. start_distillation begin with local materials 4. continue_distillation × N answer follow-ups until done 5. get_expert_detail review the generated configuration 6. set_billing pricing and trial 7. test_expert one real run (this bills) 8. publish_expert submit for review 9. get_submission_status track the outcome ``` ## Boundaries and failure modes | Symptom | Cause | Fix | | -------------------------------------------------- | ---------------------------------------- | ----------------------------------------- | | Tool reports unauthenticated | Missing Bearer token or expired cookie | Sign in again and reconnect MCP | | `install_skill` succeeds but nothing installs | The tool is a stub | Use Studio or the REST endpoint | | Agent becomes free after `METERED` | Enum naming drift | Use `ONE_TIME` + `price` to charge | | `publish_expert` succeeds but nothing changes live | It submits for review | Track with `get_submission_status` | | Distillation tools hang for a long time | They consume an SSE stream synchronously | Raise the IDE's MCP timeout | | `test_expert` consumed credits | It really executes | Use `open_expert_chat` for a free preview | ## Verify your setup Have the model call `list_my_experts` once. A returned list proves both auth and transport work; an unauthenticated error means the token never arrived. If you are unsure whether tools were discovered, check that the profy entry in your IDE's MCP panel lists 22 tools. A wrong count usually means the transport was configured as stdio instead of streamable-http. ## Related pages The equivalent REST surface How distillation works How to read evolution suggestions Real price and trial constraints Verified 2026-08-11. Sources: `services/core/src/routes/mcp/tools/index.ts` (22 `registerTool` calls), `services/core/src/routes/mcp/index.ts` (transport), `services/core/src/routes/mcp/auth.ts` (auth).