# doubleoh-mcp [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/cheedli/doubleoh  
**GitHub Stars:** 1  
**npm Downloads (last month):** 1716  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/doubleoh-mcp

## Description
The fix desk for AI agents: human fixes become skills the whole fleet reuses.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "doubleoh-mcp": {
    "command": "npx",
    "args": ["-y","doubleoh-mcp"],
    "env": {
      "DOUBLEOH_API_KEY": "",
      "DOUBLEOH_ADMIN_EMAIL": "",
      "DOUBLEOH_ADMIN_PASSWORD": ""
    }
  }
}
```

**Requires environment variables:** `DOUBLEOH_API_KEY`, `DOUBLEOH_ADMIN_EMAIL`, `DOUBLEOH_ADMIN_PASSWORD` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What doubleoh-mcp does

The doubleoh-mcp MCP server connects an MCP-capable agent to a self-hosted DoubleOh deployment. It exposes the main workflow for finding and learning procedures: retrieve skills relevant to a task and site, ask for help when the agent is blocked, and report whether a skill succeeded.

A successful fix is turned into a bounded skill rather than a raw interaction transcript. The resulting procedure contains numbered steps, a condition describing when the task is complete, and an authority statement. That statement identifies actions that still require a person and actions the agent must not take.

Skills are scoped to the customer and the site where they were learned. They use the documented Agent Skills format and can retire when they stop working. Each intervention also has an oversight record containing details such as who requested the fix, who intervened, the authority involved, and what the fleet learned.

## How it works

An agent first calls `skillsFor(task, { url })` to check whether DoubleOh already has a procedure for the current task and site. If the same obstacle is known, the learned skill can be returned without paging a person. Otherwise, `requestFix({ url, task })` creates a fix request and provides a link for a person to take control of the agent’s live browser or desktop session.

The person completes the blocked step. Typed text is not recorded, so credentials and one-time codes remain with the person. A vision model then turns the recorded interaction into a procedure with explicit boundaries. The agent can call `reportSkill(skill.name, worked)` afterward so the skill library reflects whether the procedure was effective.

The doubleoh-mcp MCP server is intended for MCP-capable agents that need these three operations without adding a framework-specific adapter. The repository also mentions adapters for several agent frameworks, but those are separate integration paths.

## Setup and configuration

The MCP package is published for use with `npx doubleoh-mcp`. It connects to a DoubleOh deployment, such as one started from the repository’s Docker Compose setup. Self-hosting requires Docker, at least 4 GB of memory, and an Anthropic key for the skill compiler. The initial image pull is approximately 5 GB.

For a local deployment, the documented startup flow copies `.env.example` to `.env`, starts the containers, and opens the application at `http://localhost:3001`. The administrator signs in with `DOUBLEOH_ADMIN_EMAIL` and `DOUBLEOH_ADMIN_PASSWORD`. The SDK example uses `DOUBLEOH_API_KEY` and a configurable `baseUrl`; MCP client configuration should provide equivalent access to the DoubleOh instance.

The repository states that configured model calls are the only traffic that leaves the local network. Deployment and configuration details, including external database and platform options, are maintained in the project documentation.

## Tools and capabilities

The doubleoh-mcp MCP server provides these core operations:

- `skillsFor`: find existing skills for a task and URL.
- `requestFix`: reuse a known fix or request human intervention for an unresolved obstacle.
- `reportSkill`: record whether a named skill worked.

The workflow supports public web targets by default. A native runtime can be used for an ERP, Windows application, or staging portal on a machine that can access it. That runtime makes one outbound connection and does not listen for inbound connections.

## Limitations and notes

Public, loopback, private, and metadata addresses are restricted by default. A host must be allowlisted for use with a custom runtime. The material does not describe the MCP server’s complete transport or client-specific configuration, so connection details may depend on the MCP client and the DoubleOh deployment.

Human intervention remains necessary for actions outside a skill’s authority boundary. Credentials and one-time codes are not recorded, and the system retains an audit trail of interventions. The repository uses the Sustainable Use License for the self-hosted application, while the MCP package is listed as MIT-licensed. The Sustainable Use terms do not allow hosting the software as a service for others without a commercial license.

_Full upstream README: https://allmcps.com/mcp/doubleoh-mcp/readme_