jirade

MCP server that gives Claude Code tools for GitHub and dbt CI on Databricks.

v0.10.0: Jira and Confluence tools were removed. Use the Atlassian MCP server in Claude Code alongside jirade — add it to .claude.json and authenticate with /mcp. No Atlassian OAuth app, scopes, or JIRADE_JIRA_OAUTH_* env vars are needed any more.

Integration dependencies

Integration	Required for	Auth
GitHub	PR tools, dbt CI (diff reports), advisors	gh auth login (auto-detected) or JIRADE_GITHUB_TOKEN
Databricks	dbt CI, UAT reports, airflow tests	Databricks CLI OAuth (default) or PAT; needs JIRADE_DATABRICKS_HOST, _HTTP_PATH, _CI_CATALOG
git (local)	change detection for CI	none
Atlassian	— none in jirade —	Atlassian MCP server in .claude.json + /mcp (see below)
Anthropic API (optional)	advisor auto-suggestions for unclassified tables	ANTHROPIC_API_KEY — without it, advisors report the gap for the calling agent to fill

What it does

jirade exposes tools via the Model Context Protocol that let Claude Code:

• Monitor GitHub PRs -- list PRs, check CI status, watch until checks pass
• Run dbt CI on Databricks -- build models in isolated schemas, compare against production using metadata-only queries, post diff reports to PRs
• Generate UAT data impact reports -- run analytical aggregate queries against CI tables and post the results to the GitHub PR (the agent posts the same markdown to the Jira ticket via the Atlassian MCP server)
• Audit jirade activity -- pull a quarter's worth of PR data (plus the JQL queries for the agent to run via the Atlassian MCP server) so the agent can write a funnel-style activity report
• Analyze dbt deprecation impact -- find downstream models affected by deprecating a table or column

No raw data is ever exposed. The Databricks client enforces a strict SQL whitelist -- only aggregated metadata queries (counts, schemas, NULLs, distributions) are allowed.

Setup

Install

# From source (recommended for development)
git clone https://github.com/djayatillake/jirade.git
cd jirade
poetry install

# Or via pipx
pipx install git+https://github.com/djayatillake/jirade.git

Configure Claude Code

Add jirade as an MCP server in your Claude Code settings (~/.claude/settings.json or project .claude/settings.json):

{
  "mcpServers": {
    "jirade": {
      "command": "jirade-mcp",
      "env": {}
    }
  }
}

If installed via poetry (not pipx), use the full path:

{
  "mcpServers": {
    "jirade": {
      "command": "/path/to/jirade/.venv/bin/jirade-mcp",
      "env": {}
    }
  }
}

Environment variables

Atlassian (Jira + Confluence): no jirade configuration. Add the Atlassian MCP server to ~/.claude.json and authenticate it with /mcp — it provides JQL/CQL search, issue reads/comments/transitions, and Confluence page create/update, acting as the authenticated user.

"mcpServers": {
  "atlassian": {
    "type": "http",
    "url": "https://mcp.atlassian.com/v1/mcp/authv2"
  }
}

Restart Claude Code after editing — server lists are built at startup — then run /mcp. Pass cloudId as your site hostname (e.g. your-org.atlassian.net). The older /v1/sse endpoint stopped being supported on 2026-06-30. The claude.ai "Atlassian Rovo" connector exposes the same tool names, but only reaches surfaces that inject claude.ai connectors; the .claude.json entry also works in the terminal CLI.

Required for GitHub tools:

# Option 1: gh CLI (recommended -- auto-detected, no env var needed)
gh auth login

# Option 2: manual token
JIRADE_GITHUB_TOKEN="ghp_..."

Required for dbt CI tools:

JIRADE_DATABRICKS_HOST="dbc-xxxxx.cloud.databricks.com"
JIRADE_DATABRICKS_HTTP_PATH="/sql/1.0/warehouses/abc123"
JIRADE_DATABRICKS_AUTH_TYPE="oauth"           # default, uses Databricks CLI creds
JIRADE_DATABRICKS_CI_CATALOG="development_yourname_metadata"  # catalog for CI schemas

Optional:

Variable	Default	Description
JIRADE_DATABRICKS_TOKEN	--	Databricks PAT (if auth_type=token)
JIRADE_DATABRICKS_CATALOG	--	Default catalog for production lookups
JIRADE_DBT_EVENT_TIME_LOOKBACK_DAYS	3	Days of data for incremental CI builds
JIRADE_DBT_CI_SCHEMA_PREFIX	jirade_ci	Prefix for CI schema names
JIRADE_LOG_LEVEL	INFO	Logging level
ANTHROPIC_API_KEY	--	Optional: advisor auto-suggestions only (Claude Code is the harness)
JIRADE_CLAUDE_MODEL	claude-opus-4-5-20251101	Model for the optional direct API calls above

Authenticate

jirade auth login                    # all services (GitHub + Databricks)
jirade auth login --service=databricks  # validate Databricks connection
jirade health                        # verify everything works

MCP tools

These tools are available to Claude Code when jirade is configured as an MCP server.

GitHub

Tool	Description
jirade_list_prs	List PRs for a repository
jirade_get_pr	Get PR details including reviews and comments
jirade_get_ci_status	Get CI check status for a PR
jirade_watch_pr	Poll CI status until all checks pass or fail (default: 30s interval, 30min timeout)

dbt

Tool	Description
jirade_run_dbt_ci	Build models on Databricks in isolated CI schemas, compare against prod, post report to PR
jirade_analyze_deprecation	Find downstream models affected by deprecating a table or column
jirade_generate_schema_docs	Read model + upstream SQL from manifest for writing lineage-aware schema descriptions
jirade_cleanup_ci	Drop CI schemas after a PR is merged
jirade_uat_report	Run analytical aggregate queries against CI tables and post the report to the PR (returns the markdown for the agent to post to Jira via the Atlassian MCP server)
jirade_test_airflow_dag	Validate an Airflow DAG's SQL by running it in a CI schema and checking idempotency

Activity audits

Tool	Description
jirade_activity_report	Pull the PR data needed for a jirade activity audit, plus the JQL queries for the agent to run via the Atlassian MCP server. Surfaces self-authored PRs, other-author PRs the user reviewed or committed to, and other users running jirade tools (cross-user discovery). Returns structured data — agent writes the narrative each run. Designed for weekly/monthly cadence.

How dbt CI works

jirade_run_dbt_ci is the main CI tool. When invoked:

1. Checks out the PR branch
2. Detects changed models and seeds from the git diff
3. Loads changed seeds via dbt seed — a changed seed that is not loaded into the CI schema aborts the run (otherwise downstream models would read the production seed and report "no changes")
4. Builds modified models +1 dependents in isolated schemas (jirade_ci_{pr_number}_{catalog}_{schema})
5. Uses --defer --state --favor-state so upstream models resolve to production
6. Compares all built models (changed + downstream) against production using metadata queries
7. For incremental/microbatch models with event_time, date-filters comparisons to the CI lookback window
8. Skips comparison for downstream models whose upstream is time-limited (CI data inherently incomplete)
9. Posts a diff report to the PR

dbt run and dbt test are separate steps so test failures don't skip downstream model builds. If some models fail but others succeed, you still get a report with a "Build Failures" section.

Only results written by this CI run count: target/run_results.json is cleared before each dbt step, so a run that dies before executing anything (expired Databricks token, parse error, empty selection) fails with the dbt output tail rather than reporting whatever an earlier dbt command left behind.

CI tables persist after the run for manual inspection. Use jirade_cleanup_ci after the PR is merged.

What the agent can see on Databricks

The DatabricksMetadataClient enforces a strict regex whitelist on every SQL query:

• DESCRIBE TABLE, SHOW COLUMNS -- column names and types
• SELECT COUNT(*) -- row counts (with optional WHERE for date filtering)
• SELECT COUNT(*) WHERE col IS NULL -- null counts
• SELECT COUNT(DISTINCT col) -- cardinality
• SELECT col, COUNT(*) GROUP BY col -- value distributions
• SELECT MIN/MAX(col) -- numeric ranges
• CREATE/DROP SCHEMA, DROP TABLE -- CI lifecycle

Everything else is rejected. No SELECT *, no raw rows, no freeform SQL.

Macros required in your dbt project

Your dbt project needs generate_schema_name and generate_database_name macros that check the DBT_JIRADE_CI environment variable to redirect models into CI schemas.

CLI

jirade list-prs --config .jirade.yaml                # List GitHub PRs
jirade health                                         # Test all connections
jirade auth status                                    # Show auth status
jirade config validate .jirade.yaml                   # Validate config
jirade env check --config .jirade.yaml               # Check environment
jirade learn status                                   # Show pending learnings

Repository config

The CLI requires a .jirade.yaml config file. Generate one with:

jirade init

License

MIT