AGENT SKILL

databricks

暂无简介

License 许可
MIT
最近更新
2026-11-04
Stars
★ 0
Forks
0

03 SKILL.md 正文

Databricks (databricks-mcp)

Read-only MCP server that runs SQL against a Databricks SQL warehouse via the Statement Execution API. FastMCP server name databricks_mcp, tool prefix databricks_.

Scope correction: this is a thin data-plane query client, not a management-plane server. There are no clusters, jobs, SQL-warehouse admin, Unity Catalog governance, workspace, DBFS, secrets, or repos tools. It queries whatever warehouse and namespace it is pointed at (default hive_metastore.default; example query SELECT count(*) FROM main.default.events).

Server at a glance

  • 3 read-only tools, zero writes. No DATABRICKS_WRITE_TOOLS_ENABLED flag exists (by design). If a write tool is ever added, gate it behind an env flag per the template's permission-tier pattern and keep the guard.
  • Two write-prevention layers: (1) list_tables/describe_table only interpolate a regex-validated dotted identifier (_validate_identifier, up to catalog.schema.table), never free text; (2) assert_read_only() in src/databricks_client.py guards run_select: strips /*...*/ and -- comments, rejects stacked statements, requires the first word in {SELECT, SHOW, DESCRIBE, WITH}, and rejects INSERT/UPDATE/DELETE/MERGE/DROP/CREATE/ALTER/TRUNCATE/GRANT/REVOKE/COPY/REPLACE as whole words. On violation it raises WriteGuardError and never touches the warehouse.
  • Transport: stdio (engineer, default) or streamable-http (alias http, for the Databricks App). Health /health, metrics /metrics, MCP endpoint /mcp/.

Auth model

Single connector code path with a per-request AuthProvider header callable:

  • Local / stdio: PAT in DATABRICKS_TOKEN -> static Bearer. Token captured in a closure, never stored as an attribute (kept out of repr/logs). Local-dev only.
  • Databricks App (prod): no PAT. The runtime injects service-principal OAuth M2M creds (DATABRICKS_CLIENT_ID/DATABRICKS_CLIENT_SECRET); main._resolve_auth builds databricks.sdk.core.Config(host=...) and passes cfg.authenticate as the provider, which returns a fresh Bearer per request and refreshes the ~hourly token. Header fetched per request, never cached on the pooled httpx client. (Fixed in commit 7a4b4c9: the App uses OAuth M2M, not a PAT.)
  • Claude Code attach: scripts/mcp_app_launcher.ps1 mints a fresh U2M token each spawn via databricks auth token --profile <your-oauth-profile> and bridges stdio to the App /mcp with npx mcp-remote.

API surface: only the SQL Statement Execution API v2.0 (POST /api/2.0/sql/statements, GET /api/2.0/sql/statements/{id}). No Jobs 2.1, Unity Catalog REST, or SCIM. Host shape https://adb-<workspace-id>.<n>.azuredatabricks.net.

Env vars

Var Default / example Purpose
DATABRICKS_HOST https://adb-<workspace-id>.<n>.azuredatabricks.net Workspace URL (not a secret)
DATABRICKS_TOKEN unset in prod PAT, local-dev only. Never log/commit
DATABRICKS_WAREHOUSE_ID <warehouse-id> Target SQL warehouse
DATABRICKS_CATALOG / DATABRICKS_SCHEMA hive_metastore / default Default namespace
DATABRICKS_APP_PORT injected by App Bind port (takes precedence over MCP_PORT)
DATABRICKS_CLIENT_ID / DATABRICKS_CLIENT_SECRET injected by App Service-principal OAuth M2M (consumed by the Databricks SDK)
MCP_TRANSPORT / MCP_PORT / MCP_HOST / MCP_SERVER_NAME / LOG_FORMAT / LOG_LEVEL stdio / 8080 / 0.0.0.0 / databricks_mcp / text / INFO Fleet transport + logging

Tool inventory

All read-only (readOnlyHint:True). All take response_format: markdown|json (default markdown). Input models forbid extra keys and strip whitespace.

Tool Purpose SQL issued Key params
databricks_list_tables List tables in a schema SHOW TABLES IN <schema> schema_name (default hive_metastore.default, validated dotted identifier)
databricks_describe_table Describe a table's columns DESCRIBE TABLE <table> table (required, validated dotted identifier)
databricks_run_select Run one arbitrary read-only statement the statement, post read-only guard statement (1-10000 chars, non-blank)

API specifics

  • DatabricksConnector.execute() posts to /api/2.0/sql/statements with {warehouse_id, statement, catalog, schema, wait_timeout:"25s", format:"JSON_ARRAY"}, then polls PENDING/RUNNING (default 5 attempts x 2s, httpx timeout 60s). Any non-SUCCEEDED terminal state raises DatabricksError with the status.error.message.
  • Result shape: columns from manifest.schema.columns[].name, rows from result.data_array (empty list if absent). Returned to tools as {"columns": [...], "rows": [[...]]}.
  • Poll budget is bounded (~25s inline wait + ~10s polling). Queries that exceed it raise DatabricksError rather than waiting longer. This is a design limit, not a bug: for long queries, narrow the query.
  • Default namespace hive_metastore.default is the legacy Hive metastore, not a Unity Catalog catalog. Pass a fully-qualified catalog.schema.table to reach elsewhere.

Deployment

Primary prod target is a Databricks App (app.yaml runs python src/main.py with MCP_TRANSPORT=streamable-http, LOG_FORMAT=json; the runtime injects DATABRICKS_APP_PORT + OAuth M2M creds). Deploy with databricks apps create mcp-databricks (the App resource name MUST start with mcp- or the AI Playground won't list it), databricks sync, databricks apps deploy. The caller IP must be on the workspace IP access list or the App returns 403. app.yaml + requirements.txt (kept in sync with pyproject by hand) drive the App build. Also runnable as stdio (engineer mode) or a Streamable-HTTP container.

Fleet conventions (adding a tool)

Scaffolded from mcp-template-build, conforms to the fleet conformance contract. Same shape as [[microsoft-graph]], [[sonarqube]], [[cribl]].

  1. Subclass BaseModule (or extend TablesModule) in src/custom_components/; discover_modules() instantiates each subclass at startup. Register under fn.__name__, which MUST start with databricks_.
  2. Input is a single Pydantic BaseModel (ConfigDict(extra="forbid"), Field(...)); set all four annotations. Validate any identifier interpolated into SQL with _validate_identifier; never interpolate raw user free-text into a statement.
  3. respx-mocked tests/test_<area>.py (mock conn.execute); no live warehouse in unit tests.
  4. Pre-push gate: pytest -v + ruff check src tests + mypy src (strict) + check_pydantic_inputs.py src + check_no_print.py src (stdout is the JSON-RPC channel; log to stderr) + check_jwt_allowlist.py.

Python >=3.12,<4.0, Poetry for dev (pip + requirements.txt for the App/Docker build). Do not run the container conformance Docker gate locally; stdio release is githooks only.

04 同仓库更多 Skills

justin-pitt
justin-pitt/pitt-skills

vibesec

This skill helps Claude write secure web applications. Use this when working on any web application or when a user requests a scan or audit to ensure security best practices are followed.

★ 0
justin-pitt
justin-pitt/pitt-skills

ui-ux-guide

Modern, clean UI/UX guidance + review skill. Use when you need actionable UX/UI recommendations, design principles, or a design review checklist for new features or existing systems (web/app). Focus on CRAP (Contrast/Repetition/Alignment/Proximity) plus task-first UX, information architecture, feedback & system status, consistency, affordances, error prevention/recovery, and cognitive load. Enforce a modern minimal style (clean, spacious, typography-led), reduce unnecessary copy, forbid emoji as icons, and recommend intuitive refined icons from a consistent icon set.

★ 0
justin-pitt
justin-pitt/pitt-skills

tufin

暂无简介

★ 0
justin-pitt
justin-pitt/pitt-skills

tines

Tines is a no-code automation and orchestration platform commonly used as a SOAR replacement, IT automation engine, and workflow builder. Use this skill whenever the user mentions Tines, Tines stories, actions, the AI Agent action, Story copilot, Webhook actions, HTTP Request actions, Event Transform, Send to Story, Resources, Pages, Cases, Records, credentials, Connect Flows, formulas, pills (`<<slug.body.field>>` syntax), the Tines API, the Tines Terraform provider, tenants, change control, self-hosted Tines, Helm charts for Tines, Tines Tunnel, Command-over-HTTP, Tines on Azure or AWS, Tines credits and the AI credit pool, the Workbench, MCP server templates exposed by Tines, or any automation or workflow build inside Tines. Also trigger when the user asks about migrating from Cortex XSOAR / XSIAM or other SOAR platforms to Tines, designing playbook or story patterns, integrating security tools through Tines, building forms with Pages, running case management in Tines, or is debugging a Tines workflow, even if they don't name a specific Tines concept.

★ 0
justin-pitt
justin-pitt/pitt-skills

tines-3b

Tines 3B is a separate product from Tines (Stories), not an upgrade path. Workflows are git repositories, steps are containerized code directories that pipe stdin to stdout, and an AI agent builds them from prompts. Use this skill whenever the user mentions Tines 3B, 3B spaces, 3B workflows, steps, config.toml, route_auth, route_type, webpage routes, volumes and /storage, Tines units (TUs), connectors and the credential proxy, 3B skills, watchtower, autofix, autotune, solo runs, 3B branches or push-live, the 3B REST API or MCP endpoint, GitSync, tunnels and networks, or docs.3b.tines.com. Also trigger when comparing 3B against Tines Stories, deciding which of the two a use case belongs in, or migrating between them. For classic Tines (stories, actions, pills, formulas, Cases, Records, Pages), use the `tines` skill instead.

★ 0
justin-pitt
justin-pitt/pitt-skills

threatconnect

ThreatConnect TIP and Polarity overlay expert. Use for any task involving ThreatConnect or Polarity: REST API calls (v2/v3), authentication (Token/HMAC), TQL queries, Indicators, Groups, Cases, Workflow, Tags, Security Labels, Associations, ThreatAssess scoring, CAL configuration, indicator deprecation, exclusion lists, TAXII feeds (inbound/outbound), feed metrics, dashboards, Query Cards, Threat Graph, ATT&CK Visualizer, Risk Quantifier, Polarity integrations (Core, IOC Submission, Intel Search, CAL, XSOAR), Polarity custom development (Node.js), Polarity server administration, Playbooks, Components, Workflow Playbooks, App Builder, TcEx, indicator distribution to SIEM/SOAR (XSIAM, Sentinel, Splunk, Elastic, Tines, Torq, BlinkOps), HTTP errors, HMAC signature failures, indicator formatting (ASN, CIDR, Host, URL, Registry Key), or troubleshooting. Trigger even without explicit ThreatConnect/Polarity mention if the task involves threat intelligence platform operations.

★ 0