# Symbioza [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/symbioza/claude-plugin  
**GitHub Stars:** 0  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/symbioza

## Description
Run a GPU job under a hard dollar cap and collect the files it writes.

## Claude Desktop Quick Installation
Remote MCP endpoint (confidence: high). Install path detected from listing signals. Add as a URL/SSE server in your client:

```json
"mcpServers": {
  "symbioza": {
    "url": "https://symbioza.dev/mcp"
  }
}
```

## Documentation

## What the Symbioza MCP server does

The Symbioza MCP server lets an AI agent prepare and run a containerized workload on a rented cloud GPU. Its documented use cases are fine-tuning, evaluation, and batch inference. The agent supplies a supported container image, a command, input data the remote machine can download, expected output files, and a total spending limit in US dollars.

Before anything is booked or charged, the agent can request an estimate. The response may include likely cost, a cost range, expected runtime, confidence, what the estimate covers, missing information, and available account credit. If runtime cannot be predicted, the service instead reports the hourly price rather than a total estimate.

## How it works

A typical run has four stages: estimate the workload, submit it after user approval, monitor its status, and retrieve the resulting files. The spending limit is a hard maximum. The service selects the compute machine, starts the job, and reports execution and billing separately.

Only files written to `/workspace/artifacts/` are collected. Retrieved files include download links and hashes for verification. Download links expire after 24 hours, but requesting the artifacts again produces fresh links.

## Setup and configuration

The hosted endpoint is `https://symbioza.dev/mcp`. It uses Streamable HTTP and OAuth sign-in in a browser; no API key or custom header is required. Claude Code users can install the plugin from the Symbioza marketplace, authenticate through `/mcp`, and enable the connector. The repository also documents direct setup for Cursor and VS Code.

In VS Code, add the HTTP server through **MCP: Add Server** or `.vscode/mcp.json`. In Cursor, add the endpoint to `mcp.json`, enable it in Customize, and complete browser sign-in. Users who install the Cursor plugin also receive the `run-on-symbioza` skill. Do not configure both the direct server and plugin in Cursor, because that exposes duplicate tools.

## Tools and capabilities

The Symbioza MCP server exposes nine tools:

- `estimateExecution` prepares a cost and runtime estimate.
- `submitJob` starts an approved job under its spending limit.
- `getStatus` checks execution progress.
- `inspectRun` inspects a run.
- `getArtifact` retrieves produced files.
- `listJobs` lists the account's jobs.
- `listSecrets` lists saved secrets.
- `describeDataset` describes a dataset.
- `cancelJob` cancels a job.

The plugin adds the `/symbioza:run-on-symbioza` skill, which guides the estimate-to-artifact workflow and tells the agent what information to request.

## Limitations and notes

The service accepts container images it has verified on its machines; the README currently identifies PyTorch and TensorFlow images. A stock image does not include project code, so the command must retrieve scripts and other dependencies itself. Inputs must be reachable by the remote machine, such as through URLs or signed links; local laptop files are not uploaded automatically.

A submitted job requires account credit. Estimates and job-list access work without credit, but submission is refused when the prepaid balance is insufficient. Secrets such as tokens and storage credentials should be saved in the service rather than pasted into chat. Recovery depends on job policy, available compute, remaining budget, and compatible checkpoint support.

_Full upstream README: https://allmcps.com/mcp/symbioza/readme_