Model Context Protocol
Connect your remote MCP client to EcoHash. Launch GPU workspaces, call open models or generate media using your own account and credit. Start with a free connection check.
Connect
Pick the tools you need, copy the matching configuration, then verify with a request that does not start paid work. New to MCP? API reference needs no account.
Public, read-only API reference. No account, key or sign-in needed.
server URL
https://mcp.ecohash.com/docsTerminal · Claude Code
claude mcp add --transport http ecohash-docs https://mcp.ecohash.com/docsRun this command from your project directory. If a server with this name already exists, inspect its configuration instead of adding it again.
Free connection check
Run /mcp in Claude Code and check the server is connected, then send the free check below.
verification prompt
Use EcoHash Docs MCP to find the GPU instance API endpoints, then explain the required fields for creating an instance. Do not execute the API.What success looks like
Matching endpoint paths and request fields from the API specification. No account resource is accessed.
When charges start
Free endpoint lookup. This server describes the public API and cannot execute calls against your account.
Keep your client's tool approvals enabled and review the proposed tool before running it. This page provides setup instructions; it does not connect to your account or run tools.
What you can do
Find capacity, quote a workspace and launch with a saved template or persistent storage. Get its SSH command, let your client work over SSH, save the results and terminate the instance.
Use a GPU-instance template for a workspace or a model-instance template for a serving endpoint. Available published weights mount read-only in their region. Also manage clusters and shared filesystems.
Discover model IDs, ask a vision model about an image, create embeddings, rerank search results or transcribe audio from a reachable HTTPS URL.
You can also call your own running model instance. MCP adds tools to your agent; it does not replace your client's underlying model.
Import a product photo or voice clip, discover engines, review a video quote, generate media and follow the job to its result. Images, speech and music are available too.
Keep the media ID to refresh an expired reference link. Generation produces media files; use the Studio for timeline editing.
Docs MCP searches the public OpenAPI specification and returns endpoint parameters and response definitions. It is free, needs no sign-in and cannot execute account operations.
Set up Docs MCP →Before you run
Your client runs the agent and decides how to present tool approvals. EcoHash provides the tools and bills the underlying usage on your account.
Terminating releases the GPU and deletes temporary instance data. It is not a pause you can resume. Save results to persistent storage or copy them out first. Cloud drives and shared filesystems remain and continue billing until deleted; deleting them destroys their data.
Compute Read-only prevents GPU launches, resource changes and deletion. It does not make the entire key read-only: keys created in the console can also have inference access, which permits paid model and media calls. Review each permission, and use only the servers you need.
GPU spending and destructive tools require a confirmation parameter. Your client supplies it; it is not a separate EcoHash approval dialog. Keep client approvals enabled, quote launches and videos before committing, and review the tool and amount. Inference and media tools can spend credit when called.
A no-end-time instance normally holds the first hour of credit and continues billing until terminated or account limits stop it. A fixed-duration launch normally holds the run's cost up front. Review the returned quote and launch details. Persistent storage is billed separately until deleted.
Disconnecting an MCP client or revoking its key does not terminate running resources. Video generation is asynchronous: keep the job ID and check its progress instead of submitting it again. Completed download links may expire; check the job again for a fresh link. Terminate GPUs separately when finished.
Use a client that supports remote Streamable HTTP; the server URL is not a webpage, so opening it in a browser is not a connection test. Check credentials, environment variables and permissions in your client's MCP status. A Compute permission error needs the correct access, not repeated launches. Try Docs MCP without credentials to check basic connectivity. Client formats are documented in the linked guides; support also depends on your version and organization settings.