cdash-mcp
Browse, inspect, and triage Kitware CDash CI/CD data through read-only MCP tools.
Dashboards & project overview:
get_dashboardandget_project_overviewshow build groups, build/config/compile/test counts, and aggregate coverage/analysis statistics for a project on a given date.Build inspection:
get_build_detailsreturns counts, durations, site, compiler and revision;get_build_errorslists compiler errors or warnings with source locations;get_configure_outputshows the CMake configure command and log (with offset/limit slicing).Source changes:
get_build_updateshows VCS commits (revision and update status) associated with a build.Test triage:
get_failing_testsfinds non-passing tests across builds for a project/date;get_build_testslists a build's tests with status filters;get_test_summarytracks a test's pass/fail history across builds to spot flakiness.Coverage:
get_coverage_comparisoncompares coverage across builds (or a specificbuild_id) to detect regressions, with paginated file results.Dynamic analysis:
get_dynamic_analysissurfaces Valgrind/sanitizer-style defect entries for a build.Pagination & filtering: most list tools support
limit/offset, status filters, name filters, and date ranges.Authentication checks: works against public or private CDash instances via
CDASH_URL/CDASH_TOKEN, with all tools read-only.Note: the README advertises 26 tools (e.g. build triage, test details/images/history, artifacts, commands, notes), but this schema exposes the 11 listed above.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@cdash-mcp show failing tests for the thor project"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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-mcpOpenAI 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 listThe 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-mcpUse --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-mcpRelated MCP server: blacksmith-mcp
Configuration
Environment Variable | Required | Default | Description |
| No |
| CDash instance URL |
| No | — | Bearer token for authentication (required for private instances) |
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)
Tool | Purpose |
| Authentication, project access, server version and available build fields |
| Dashboard build groups and configure/compile/test counts |
| Project coverage, analysis and build-group overview |
| Build history by name, UTC date range, type or exact revision |
| Build counts, durations, site, compiler and revision |
| Summary, configure log, compiler errors and failed tests in one call |
| Errors or warnings, source locations, commands and retrievable context |
| Configure command, return value and log |
| Revision, prior revision and update status |
| Changed files, authors and commit messages |
| Non-passing test results on a CDash date, with build/test IDs |
| Per-build tests, status filters and timing statistics |
| Restored test command, measurements, status and output |
| Submitted regression images and their URLs |
| Exact-name results across builds on one CDash date |
| Exact-name history across an optional UTC build-date range |
| New/fixed test failures, statuses, runtimes and build counts |
| File line, branch and function coverage |
| Exact file's source (when permitted) and line hit counts |
| Explicit build pair, changed/added/removed files and weighted line totals |
| Compatibility alias requiring an explicit |
| Checker, command and per-type defect counts |
| Individual analysis log and defects |
| Submitted notes with retrievable text |
| Uploaded files/download URLs or submitted links |
| Submitted CMake command timings and resource measurements |
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_connectionwith your project name to verify authentication and membership.For Codex, check
~/.codex/config.toml(or.codex/config.tomlfor a trusted project), then start a new session and inspect/mcp.Make sure the
envblock is in the right config file. For Claude Code, MCP servers must be defined in~/.claude.json— putting them in~/.claude/settings.jsonwill silently ignore the env vars.After changing config, restart the MCP server (
/mcpin 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-mcpLicense
MIT
Available Tools
11 toolsget_build_detailsB
Get detailed information about a specific build, including configure/compile/test summaries.
Args:
build_id: The CDash build ID.
Inputs: build_id (required). Returns: result
get_build_errorsA
View compiler errors or warnings for a build, with source file and line info.
Args:
build_id: The CDash build ID.
warnings: If True, show warnings instead of errors.
limit: Maximum number of errors to return (default 30, max 200).
offset: Number of errors to skip (default 0). Use for pagination.
Inputs: build_id (required), limit, offset, warnings. Returns: result
get_build_testsB
List tests for a specific build, optionally filtered by status.
Args:
build_id: The CDash build ID.
status_filter: Optional filter: "passed", "failed", or "notrun".
limit: Maximum number of tests to return (default 50, max 200).
offset: Number of tests to skip (default 0). Use for pagination.
Inputs: build_id (required), limit, offset, status_filter. Returns: result
get_build_updateB
View source code changes (VCS commits) associated with a build.
Args:
build_id: The CDash build ID.
Inputs: build_id (required). Returns: result
get_configure_outputA
View CMake configure command and output for a build.
Args:
build_id: The CDash build ID.
output_offset: Character offset into the configure output (default 0).
output_limit: Maximum characters of output to return (default 34816 = 34 KB).
Set to 0 for no limit.
Inputs: build_id (required), output_limit, output_offset. Returns: result
get_coverage_comparisonA
Compare code coverage across builds for a project. Useful for detecting coverage regressions.
Args:
project: CDash project name (e.g. "PublicDashboard").
date: Optional date (YYYY-MM-DD). Defaults to today.
build_id: Optional build ID to get coverage for a specific build.
Recommended: provide a build_id from the dashboard for reliable results.
Without build_id, uses cross-build comparison (only works for Nightly builds).
limit: Maximum number of files to return (default 50, max 200).
offset: Number of files to skip (default 0). Use for pagination.
Inputs: project (required), date, limit, offset, +1 more. Returns: result
get_dashboardA
Get the CDash dashboard for a project, showing build groups and status.
Args:
project: CDash project name (e.g. "PublicDashboard").
date: Optional date (YYYY-MM-DD). Defaults to today.
Inputs: project (required), date. Returns: result
get_dynamic_analysisA
Get dynamic analysis results (e.g. Valgrind, sanitizers) for a build.
Args:
build_id: The CDash build ID.
limit: Maximum number of defect entries to return (default 50, max 200).
offset: Number of defect entries to skip (default 0). Use for pagination.
Inputs: build_id (required), limit, offset. Returns: result
get_failing_testsA
Find non-passing tests across all builds for a project. Most useful for CI triage.
Args:
project: CDash project name (e.g. "PublicDashboard").
date: Optional date (YYYY-MM-DD). Defaults to today.
test_name: Optional filter to match test names containing this string.
limit: Maximum number of tests to return (default 50, max 200).
offset: Number of tests to skip (default 0). Use for pagination.
Inputs: project (required), date, limit, offset, +1 more. Returns: result
get_project_overviewA
Get project overview with aggregate build/test/coverage statistics.
Args:
project: CDash project name (e.g. "PublicDashboard").
date: Optional date (YYYY-MM-DD). Defaults to today.
Inputs: project (required), date. Returns: result
get_test_summaryA
Get summary of a test across builds — shows pass/fail history to detect flaky tests.
Args:
project: CDash project name (e.g. "PublicDashboard").
test_name: Exact name of the test.
date: Optional date (YYYY-MM-DD). Defaults to today.
limit: Maximum number of builds to return (default 50, max 200).
offset: Number of builds to skip (default 0). Use for pagination.
Inputs: project (required), test_name (required), date, limit, +1 more. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
Most tools are clearly separated by resource type, such as build details, errors, tests, coverage, and dynamic analysis. However, get_dashboard and get_project_overview both provide project-level aggregate status, and get_failing_tests overlaps somewhat with get_build_tests with a failure filter, creating minor ambiguity.
All tools follow the get_<resource> pattern consistently, making the set highly predictable. The names clearly indicate both the action (get) and the target (dashboard, build, coverage, tests, etc.).
With 11 tools, the count is well within the ideal range for a domain-specific read-only CDash server. Each tool covers a meaningful slice of the dashboard/build/test/coverage surface without unnecessary duplication.
The tool set covers the main CDash read workflows: dashboards, build details, test results, errors, coverage, dynamic analysis, configure output, and VCS update info. The main gap is project discovery—there is no way to list available projects, so agents must already know the project name.
Maintenance
Related MCP Connectors
Test analytics for AI agents: test history, flaky tests, failure clusters, coverage.
Read your runners.io organisation's jobs, logs, test results, cache and spend from an AI agent.
Direct access to Cypress tests results and accessibility reports in your AI workflow.
Ask about your Kodem Security data and act on findings in plain language, with your permissions.
Related MCP Servers
AlicenseNot gradedqualityBmaintenanceEnables access to Codacy's code quality platform through natural language, providing repository management, security analysis, pull request reviews, and local CLI-based code analysis. Supports comprehensive code quality monitoring including issues, coverage, security vulnerabilities, and technical debt assessment across organizations and repositories.3,691 npm58MIT- AlicenseBqualityDmaintenanceEnables interaction with Blacksmith CI analytics to query workflow runs, jobs, test results, and usage metrics. It provides detailed access to CI/CD data including job logs and billing information directly through Claude.37330 npm6MIT
- AlicenseNot gradedqualityDmaintenanceEnables management of TeamCity CI/CD operations through natural language commands, including builds, projects, tests, artifacts, and more.16 npmMIT
- AlicenseAqualityCmaintenanceEnables interaction with SonarCloud projects, issues, quality gates, and security hotspots through natural language.1555 PyPIMIT