# cuda4cpu (oci · ghcr.io/javier-cabezas/cuda4cpu)

Compile, run and check CUDA and HIP kernels without a GPU: wrong results, races, memory errors

- Trust score: 40/100 (low)
- Change this week: +3
- Registry status: active
- Liveness: live
- Owner verified: no
- Last scored: 2026-11-05

## Components

- oci · `ghcr.io/javier-cabezas/cuda4cpu`: 40/100 (this document), [markdown](https://verifymcp.io/servers/javier-cabezas-cuda4cpu/ghcr-io-javier-cabezas-cuda4cpu.md), [page](https://verifymcp.io/servers/javier-cabezas-cuda4cpu/ghcr-io-javier-cabezas-cuda4cpu)

## Channel facts

- Registry: `oci`
- Package: `ghcr.io/javier-cabezas/cuda4cpu`
- Version: `0.5.0`
- Transport: `stdio`

## Trust breakdown

How this component scores in each security and reliability category. Every signal is checked automatically from public evidence about the published package, including repeated runs of it in an isolated sandbox, and we only credit what we can confirm. Scores are 0–100 per category. Scoring method: https://verifymcp.io/docs/scoring (what has changed: https://verifymcp.io/docs/scoring/changelog)

Scored 2026-11-05.

- **Supply Chain Security**: 0/100
  - No malware scan is available for this kind of package: the supply-chain vendors we use do not cover it. This is a permanent gap in our coverage, not a finding about the package.
  - Known CVEs could not be checked: this artifact ships no SBOM or dependency manifest, so there is no dependency list to read.
  - Install-script risk not yet assessed.
  - Dependency health could not be checked: this artifact ships no SBOM or dependency manifest, so there is no dependency list to read.
- **Provenance & Transparency**: 45/100
  - Source repository is publicly reachable at the declared URL.
  - Provenance check failed: no build-provenance attestation is published.
  - Clear OSI-approved license (Apache-2.0).
  - Actively maintained (last published 1 days ago).
  - Disclosure check failed: no security disclosure policy was found in the source repository.
- **Schema Quality & AI Usability**: 56/100
  - AI-judged instruction clarity (good).
  - Context-footprint check failed: tool/resource definitions use about 1605 tokens (~321/item across 5 items; 5 tools + 0 resources), over budget; trim descriptions and params.
  - Usage-examples check failed: none of the tools include examples.
- **Stability & Change Management**: 33/100
  - Stability observed for 10 of 30 days with no destabilising changes; credit accrues until the full window elapses.
- **Tool Coverage**: 100/100
  - 100% of tools have a non-trivial description (not blank, and not just the tool's name).
  - 100% of tool parameters carry a description.
- **Tool Safety**: 100/100
  - No prompt-injection markers were found in the server instructions, tool names or descriptions we captured.
  - We read all 5 captured tool definition(s), and no name or description among them implies an irreversible operation.
  - An AI judge read all 6 captured unit(s) of tool text and found none that tries to manipulate the model reading it.
- **Capabilities**: 60/100
  - Spec-recency check failed: implements MCP spec 2025-06-18; the latest is 2026-07-28.

**Unverified: 1 category.** A category scored 0 because we could not verify it: a data source with nothing on this package, evidence we could not reach, or a check we could not run. We only credit what we can confirm.

## Install

### How do I install the cuda4cpu MCP server?

cuda4cpu runs locally as a container image, launched with docker run --rm -i ghcr.io/javier-cabezas/cuda4cpu:0.5.0. Ready-made configuration for Claude, Cursor, VS Code, Codex and 3 more is on this page, copied from each client's own documentation.

### Claude

```bash
claude mcp add javier-cabezas-cuda4cpu -- docker run --rm -i ghcr.io/javier-cabezas/cuda4cpu:0.5.0
```

### Cursor

```json
{
  "mcpServers": {
    "javier-cabezas-cuda4cpu": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "ghcr.io/javier-cabezas/cuda4cpu:0.5.0"
      ]
    }
  }
}
```

### VS Code

```json
{
  "servers": {
    "javier-cabezas-cuda4cpu": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "ghcr.io/javier-cabezas/cuda4cpu:0.5.0"
      ]
    }
  }
}
```

### Codex

```bash
codex mcp add javier-cabezas-cuda4cpu -- docker run --rm -i ghcr.io/javier-cabezas/cuda4cpu:0.5.0
```

### opencode

```json
{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "javier-cabezas-cuda4cpu": {
      "type": "local",
      "command": [
        "docker",
        "run",
        "--rm",
        "-i",
        "ghcr.io/javier-cabezas/cuda4cpu:0.5.0"
      ],
      "enabled": true
    }
  }
}
```

### Hermes

```yaml
mcp_servers:
  javier-cabezas-cuda4cpu:
    command: "docker"
    args: ["run", "--rm", "-i", "ghcr.io/javier-cabezas/cuda4cpu:0.5.0"]
```

### Netclaw

```json
{
  "McpServers": {
    "javier-cabezas-cuda4cpu": {
      "Transport": "stdio",
      "Command": "docker",
      "Arguments": [
        "run",
        "--rm",
        "-i",
        "ghcr.io/javier-cabezas/cuda4cpu:0.5.0"
      ]
    }
  }
}
```

### Other

```json
{
  "mcpServers": {
    "javier-cabezas-cuda4cpu": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "ghcr.io/javier-cabezas/cuda4cpu:0.5.0"
      ]
    }
  }
}
```

## Changelog

Every change recorded for this component, newest first. Days that predate change tracking, or that we cannot explain, say so: "we were watching and nothing happened" and "we were not watching" are different claims.

### 2026-11-04 (score 40, +1)

- [functional] Package version: 0.4.1 → 0.5.0

### 2026-11-02 (score 39, +1)

No change was recorded against any check on this day. Stability & Change Management went from 20 to 23. That category is still filling its 30-day observation window: 6 days of observed history at the previous scan, 7 at this one. The score rises as the window fills, whether or not the server changes.

### 2026-10-31 (score 38, +1)

No change was recorded against any check on this day. Stability & Change Management went from 13 to 17. That category is still filling its 30-day observation window: 4 days of observed history at the previous scan, 5 at this one. The score rises as the window fills, whether or not the server changes.

### 2026-10-29 (score 37, +1)

No change was recorded against any check on this day. Stability & Change Management went from 7 to 10. That category is still filling its 30-day observation window: 2 days of observed history at the previous scan, 3 at this one. The score rises as the window fills, whether or not the server changes.

### 2026-10-27 (score 36, +1)

No change was recorded against any check on this day. Stability & Change Management went from 0 to 3. That category is still filling its 30-day observation window: 0 days of observed history at the previous scan, 1 at this one. The score rises as the window fills, whether or not the server changes.

### 2026-10-26 (score 35)

First indexed and scored.

## MCP tools (5)

### `compile` (~203 tokens)

Compile CUDA or HIP sources with cuda4cpu, as nvcc or hipcc would: whether they build, the compiler's diagnostics at the sources' own lines, and the features the sources use that cuda4cpu does not implement (textures, inline PTX, CUDA libraries, ...).

Input parameters:

- `files` (object): Sources by file name (relative paths) and their contents: .cu or .hip files, .cpp host files, and headers they include
- `language` (string): cuda (cuda4cpu-nvcc, the default) or hip (cuda4cpu-hipcc, the default when a source is a .hip file)
- `options` (array): Compiler options, as nvcc's (-arch, -Xcompiler, -std, -I, -D, ...) or hipcc's; GPU options are accepted and ignored. Default -O2 -g.
- `paths` (array): Existing source files to compile where they are

### `run` (~277 tokens)

Compile and run a CUDA or HIP program on the CPU: its exit code, output, and the problems cuda4cpu reports (divergent barriers, copies in the wrong direction). Use check to look for bugs.

Input parameters:

- `args` (array): The program's arguments
- `environment` (object): More environment variables for the program
- `files` (object): Sources by file name (relative paths) and their contents: .cu or .hip files, .cpp host files, and headers they include
- `language` (string): cuda (cuda4cpu-nvcc, the default) or hip (cuda4cpu-hipcc, the default when a source is a .hip file)
- `options` (array): Compiler options, as nvcc's (-arch, -Xcompiler, -std, -I, -D, ...) or hipcc's; GPU options are accepted and ignored. Default -O2 -g.
- `paths` (array): Existing source files to compile where they are
- `stdin` (string): Input for the program
- `timeout_seconds` (number): Time limit for each run (default 60, at most 600)
- `working_directory` (string): Where the program runs: by default, the directory of the last path given, or the directory of the inline files

### `check` (~478 tokens)

Look for bugs in a CUDA or HIP program without a GPU. Runs it with the threads of each block in several orders (forward, reverse, random with several seeds, warp orders): a result that changes with the order means a missing __syncthreads() or __syncwarp(), or a race. Also reports cuda4cpu's checks (divergent barriers and warp operations, copies in the wrong direction, allocations never freed), runs once with new allocations filled with 0xff bytes (a changed result means reads of memory never written), and runs a build with AddressSanitizer and UBSan (out-of-bounds and use-after-free accesses, undefined behavior), with source lines. The program should check its own results or print them; output that varies between identical runs (timings) is compared by exit code only, unless ignore_patterns removes the lines that vary.

Input parameters:

- `args` (array): The program's arguments
- `environment` (object): More environment variables for the program
- `files` (object): Sources by file name (relative paths) and their contents: .cu or .hip files, .cpp host files, and headers they include
- `ignore_patterns` (array): Regular expressions of output lines to leave out of the comparison between orders, such as timings
- `language` (string): cuda (cuda4cpu-nvcc, the default) or hip (cuda4cpu-hipcc, the default when a source is a .hip file)
- `options` (array): Compiler options, as nvcc's (-arch, -Xcompiler, -std, -I, -D, ...) or hipcc's; GPU options are accepted and ignored. Default -O2 -g.
- `paths` (array): Existing source files to compile where they are
- `sanitizers` (boolean): Also run with AddressSanitizer and UBSan (default true)
- `seeds` (array): Seeds of the random orders (default 1, 2, 3)
- `stdin` (string): Input for the program
- `timeout_seconds` (number): Time limit for each run (default 60, at most 600)
- `working_directory` (string): Where the program runs: by default, the directory of the last path given, or the directory of the inline files

### `test_kernel` (~455 tokens)

Test one kernel without writing a program: give its source, the launch configuration, its arguments (scalars, and arrays with their data, given or generated), and the expected outputs, as values or as a host reference function. A generated program launches the kernel, compares each output (exactly for integers, within a tolerance for floating point) and reports the elements that differ. With check (the default), it also runs in the thread orders and with the sanitizers that check uses.

Input parameters:

- `arguments` (array, required): The kernel's arguments, in order
- `block` (integer|array, required): Threads per block: a number or [x, y, z]
- `check` (boolean): Also run in other thread orders and with the sanitizers (default true)
- `grid` (integer|array, required): Blocks: a number or [x, y, z]
- `kernel` (string, required): The source of the kernel, with what it needs (device functions, includes, constants)
- `language` (string): cuda (cuda4cpu-nvcc, the default) or hip (cuda4cpu-hipcc, the default when a source is a .hip file)
- `name` (string, required): The kernel to launch (a template with its arguments, e.g. scale<float>, works too)
- `options` (array): Compiler options, as nvcc's (-arch, -Xcompiler, -std, -I, -D, ...) or hipcc's; GPU options are accepted and ignored. Default -O2 -g.
- `reference` (string): C++ host code that defines void reference(...) with the kernel's parameters: it gets host copies of the inputs and writes the expected outputs into the output arrays (those without expected values)
- `seeds` (array): As check's
- `shared_bytes` (integer): Dynamic shared memory per block (default 0)
- `timeout_seconds` (number): Time limit for each run (default 60, at most 600)
- `tolerance` (object): For floating-point outputs: an element matches if |got - expected| <= absolute + relative * |expected| (default 1e-5 and 1e-5)

### `capabilities` (~36 tokens)

What cuda4cpu implements of CUDA and HIP, what it does not, and how its checks work: read it before relying on a feature.

## Diagnostics

Captured diagnostic sections: Provenance. The full working is on the page: https://verifymcp.io/servers/javier-cabezas-cuda4cpu/ghcr-io-javier-cabezas-cuda4cpu#diagnostics

## Score history

- 2026-11-05: 40
- 2026-11-04: 40
- 2026-11-03: 39
- 2026-11-02: 39
- 2026-11-01: 38
- 2026-10-31: 38
- 2026-10-30: 37
- 2026-10-29: 37
- 2026-10-28: 36
- 2026-10-27: 36
- 2026-10-26: 35

## Common questions

### What is the cuda4cpu MCP server?

cuda4cpu is an MCP server listed in the public MCP registry as io.github.javier-cabezas/cuda4cpu. Compile, run and check CUDA and HIP kernels without a GPU: wrong results, races, memory errors. This page covers its container image (ghcr.io/javier-cabezas/cuda4cpu).

### Is the cuda4cpu MCP server safe to use?

cuda4cpu scores 40 out of 100 on VerifyMCP. That is a record of what we were able to check automatically, not an endorsement. The category breakdown on this page shows every signal behind the number, including the ones we could not confirm.

### What tools does the cuda4cpu MCP server expose?

cuda4cpu exposes 5 tools: compile, run, check, test_kernel, capabilities. Their descriptions and schemas cost roughly 1,449 tokens of context every time the server is loaded.

### Is the cuda4cpu MCP server still maintained?

cuda4cpu is still listed as active in the MCP registry. We last reached this channel on 5 November 2026. Those dates come from our own scans of the registry and the channel itself, not from anything the publisher announced.

### What licence is the cuda4cpu MCP server under?

cuda4cpu declares the Apache-2.0 licence, which is OSI-approved. That covers the source only, and says nothing about the cost of any service it calls.

## Links

- GitHub Container Registry: https://ghcr.io/javier-cabezas/cuda4cpu
- Repository: https://github.com/javier-cabezas/cuda4cpu
- Website: https://github.com/javier-cabezas/cuda4cpu/blob/main/docs/debugging.md#checking-kernels-from-an-ai-agent-mcp
- Changelog RSS feed: https://verifymcp.io/servers/javier-cabezas-cuda4cpu/ghcr-io-javier-cabezas-cuda4cpu.xml
- Changelog JSON feed: https://verifymcp.io/servers/javier-cabezas-cuda4cpu/ghcr-io-javier-cabezas-cuda4cpu.json
- HTML version of this page: https://verifymcp.io/servers/javier-cabezas-cuda4cpu/ghcr-io-javier-cabezas-cuda4cpu