Computeledger

What the computeledger MCP server does

The computeledger MCP server gives MCP-compatible clients a local interface for creating and checking compute-usage receipts. A receipt can describe a provider, hardware type, duration, GPU-hours, estimated FLOPs, and workload type. Each recorded entry is signed with an Ed25519 private key and added to a ledger that links entries through their hashes.

The resulting records are provider-agnostic. They can be created on a laptop, cloud virtual machine, or on-premises system, then verified elsewhere without a ComputeLedger account or network connection. Verification checks both the receipt signature and the data used to produce its hash.

How it works

The server uses stdio transport: the MCP client starts computeledger mcp as a local subprocess, so no hosted endpoint is required. The Python distribution exposes this command through the computeledger-cli package. The npm distribution provides the same MCP tools through npx computeledger-cli mcp.

Keys and ledger data are stored in the ComputeLedger directory. By default, commands use the user-level location; the CLI also supports a local project directory through its --local option. The MCP tools expose the corresponding local-selection option where supported.

A receipt can be checked independently with the CLI as well as through MCP. Ledger verification detects changes to historical entries, deleted entries, or reordered entries because each receipt includes the previous receipt's hash. The npm and Python implementations use the same canonical serialization and can verify each other's receipts.

Setup and configuration

Install the Python package with:

pip install computeledger-cli

A client configuration can launch the server with uvx:

{
  "mcpServers": {
    "computeledger": {
      "command": "uvx",
      "args": ["--from", "computeledger-cli", "computeledger", "mcp"]
    }
  }
}

The npm package can also be installed globally with npm install -g computeledger-cli, then started with npx computeledger-cli mcp. Before recording usage, generate a local Ed25519 keypair with computeledger keys generate --local when using a project-local ledger. The README does not specify environment variables or required service credentials.

Tools and capabilities

The computeledger MCP server exposes four tools:

• record_usage signs usage fields and appends the resulting receipt to the ledger. It accepts provider, hardware, duration, optional GPU-hours, optional estimated FLOPs, optional workload type, and an optional local-storage flag.
• verify_receipt checks an individual receipt's signature and hash integrity.
• list_ledger returns the receipts recorded in the selected local ledger.
• verify_ledger checks every receipt and confirms that the complete hash chain remains unbroken.

The same operations are available from the CLI with structured JSON output. The separate run command can wrap a real subprocess, measure wall-clock duration, sample NVIDIA GPU utilization when nvidia-smi is available, and record the result automatically; this behavior belongs to the CLI rather than one of the four MCP tools.

Limitations and notes

The computeledger MCP server records and verifies attestations; it is not a GPU marketplace, cost dashboard, or provider integration. It does not establish that the original usage claims are truthful beyond validating the signature and subsequent ledger integrity. The README also does not describe remote ledger storage, cloud-provider API synchronization, or hosted deployment.

On systems without an NVIDIA GPU, the CLI's wrapped-job mode can still create a duration-only receipt. The MCP transport itself is local stdio, and the documented configuration assumes the client can launch a local command. Wrapped CLI commands run as argument arrays rather than shell strings, but that execution mode is not part of the MCP tool list.