CDash MCP 服务器

cdash-mcp

An MCP server for Kitware CDash — the CI/CD dashboard for projects built with CMake/CTest. Browse dashboards, find failing tests, inspect build errors, check coverage, and triage CI failures, all through natural language. Works with OpenAI Codex, Claude Desktop/Code, Cursor, and any MCP-compatible client.

Provides 26 read-only tools for navigating CDash builds, tests, coverage, and dynamic analysis.

Quick Start

Prerequisites

• Python 3.12+
• A CDash instance (defaults to my.cdash.org)

Installation

# Install from GitHub with uv (recommended)
uv tool install git+https://github.com/cbyrohl/cdash-mcp

# Or with pip
pip install git+https://github.com/cbyrohl/cdash-mcp

OpenAI Codex

Add the server to Codex with the CLI:

codex mcp add cdash \
  --env CDASH_URL=https://my.cdash.org \
  --env CDASH_TOKEN=your-token-here \
  -- uvx --from git+https://github.com/cbyrohl/cdash-mcp cdash-mcp

# Confirm that Codex stored the configuration
codex mcp list

The Codex CLI, IDE extension, and ChatGPT desktop app share this MCP configuration. Omit the CDASH_TOKEN option for public instances. Start a new Codex session after adding the server, then use /mcp to inspect its tools.

For project-scoped configuration, export your credentials and add this to .codex/config.toml in a trusted repository:

export CDASH_URL=https://my.cdash.org
export CDASH_TOKEN=your-token-here

[mcp_servers.cdash]
command = "uvx"
args = ["--from", "git+https://github.com/cbyrohl/cdash-mcp", "cdash-mcp"]
env_vars = ["CDASH_URL", "CDASH_TOKEN"]

See the Codex MCP documentation for additional configuration and tool-policy options.

Claude Code

claude mcp add cdash \
  -e CDASH_URL=https://my.cdash.org \
  -e CDASH_TOKEN=your-token-here \
  -- uvx --from git+https://github.com/cbyrohl/cdash-mcp cdash-mcp

Use --scope user for global access, --scope project to share via .mcp.json in your repo, or omit --scope for local (current project only).

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "cdash": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/cbyrohl/cdash-mcp", "cdash-mcp"],
      "env": {
        "CDASH_URL": "https://my.cdash.org",
        "CDASH_TOKEN": "your-token-here"
      }
    }
  }
}

Running from Source

git clone https://github.com/cbyrohl/cdash-mcp.git
cd cdash-mcp
uv sync

# Run the server
uv run cdash-mcp

Configuration

Note: Project names in CDash are case-sensitive (e.g. "thor" and "THOR" are different projects).

API compatibility and authentication

The modern client uses GraphQL for build inspection and supported REST endpoints for dashboards and project-wide test queries. Tested against my.cdash.org CDash 5.4. The legacy build summary, compiler error, configure, test history, source update, and coverage URL paths are no longer called. Older CDash versions are not automatically supported: use check_connection to inspect available fields, and report unsupported fields rather than treating them as empty results.

Private projects require a valid, unexpired full-access token belonging to a user with project access. Submission-only tokens cannot authenticate this client. The client first calls a REST endpoint to establish a session cookie, then uses that cookie with GraphQL. Session authentication is retried once when access is denied; persistent permission failures remain errors. No token is included in client representations or tool results.

All tools are read-only. API failures set the MCP error flag. Results are JSON objects rather than the previous Markdown summaries, and use current CDash field names. IDs from GraphQL are strings. Test-query REST results additionally include numeric build_id and test_id fields and navigable URLs.

Tools (26)

Pagination and output

Build-scoped test, compiler-diagnostic, coverage, instrumentation and dynamic-analysis lists merge the selected build's records with all immediate child builds. Each record includes build_id and subproject (ID/name or null). Parent results come first, then children in ID order, with each relation preserving its own pagination cursor. The child-build catalog is also paginated. Each result page contains records from one source build; pages can be shorter than limit at a child boundary. Continuation cursors are opaque and bound to the selected build, relation, filter, child catalog and page size; changing those invalidates the cursor. A final continuation page can be empty when remaining children have no matching records.

GraphQL lists return items and page_info. Pass page_info.endCursor as after when hasNextPage is true. Page sizes are 1–200. Keep the same limit on every continuation request; to change the page size, restart without after. Lighthouse interprets its underlying cursor using the requested page size, so the client validates this to prevent repeated or skipped records. Raw CDash cursors and cursors from older MCP versions are not accepted.

GraphQL list tools no longer accept offset; they do not fetch and discard pages to simulate skipping records. REST test queries and computed comparison results retain local limit/offset slicing. Text logs retain their independent output_offset argument.

Logs use output_offset and output_limit (default 34,816 characters). A sliced text object includes text, total_characters, offset and next_offset. Use next_offset to continue, or output_limit=0 for the complete text.

Build searches and multi-date history use inclusive UTC calendar dates; dashboard and test-summary dates follow CDash's configured dashboard day. These can differ when the nightly rollover is not midnight.

Comparisons fetch all relevant records, up to 10,000 per build, before computing changes. Larger results fail explicitly. Duplicate test names within the same subproject also fail rather than guessing which results correspond. Comparisons match tests and files by subproject ID plus name/path, so identically named records in different subprojects remain separate. Coverage totals count each submitted subproject record. A fixed failure means FAILED became PASSED; removed or skipped tests are not counted as fixes. Coverage percentages are weighted by executable line counts, and file comparisons do not compute patch coverage.

CMake instrumentation and artifacts are available only when submitted by CI. Nested command measurements expose their own pagination information; this tool currently retrieves the first 20 measurements per command.

Troubleshooting

Authentication or authorization errors:

• Verify your token is valid, unexpired and full-access in CDash under My Profile > Authentication Token.
• Run check_connection with your project name to verify authentication and membership.
• For Codex, check ~/.codex/config.toml (or .codex/config.toml for a trusted project), then start a new session and inspect /mcp.
• Make sure the env block is in the right config file. For Claude Code, MCP servers must be defined in ~/.claude.json — putting them in ~/.claude/settings.json will silently ignore the env vars.
• After changing config, restart the MCP server (/mcp in Codex or Claude Code, or restart the application).

Project not found / empty dashboard:

• CDash project names are case-sensitive. Check the exact name in your CDash instance.

Development

# Install dev dependencies
uv sync

# Run deterministic tests, including MCP and STDIO transport (no network)
uv run pytest -v

# Opt in to live smoke tests; export CDASH_TOKEN securely if needed
CDASH_URL=https://open.cdash.org CDASH_LIVE_PROJECT=PublicDashboard \
  uv run pytest -m integration -v

# Lint
uv run ruff check src/ tests/

# Run the server locally
uv run cdash-mcp

License

MIT