Galileo MCP Server Setup

Galileo MCP Server Setup

Prerequisites

Workflow Overview

┌───────────┐   ┌───────────────┐   ┌───────────────┐   ┌─────────────────┐
│ Preflight │ → │ Render/review │ → │ Apply/handoff │ → │ Validate evidence │
└───────────┘   └───────────────┘   └───────────────┘   └─────────────────┘

When to Activate

• Configuring Galileo MCP, registering Galileo with IDE/agent clients, inventorying live MCP tools, or auditing Galileo MCP product coverage. Covers Galileo API-key secret handling, self-hosted URL derivation, live MCP tool inventory and.
• Preview and review the galileo mcp server setup workflow before any live apply phase.
• Diagnose failed prerequisites, generated assets, configuration, or validation evidence.

Scope

Follow the documented read-only or render-first path whenever it is available. This skill does not imply permission to mutate live systems. Require explicit apply flags, protected credentials, and operator review for state changes.

Examples

Inspect the supported setup modes before selecting one:

bash skills/galileo-mcp-server-setup/scripts/setup.sh --help

Expected output: usage, supported modes, and required arguments are displayed without changing the target environment.

Inspect validation modes before running completion checks:

bash skills/galileo-mcp-server-setup/scripts/validate.sh --help

Expected output: offline, live, and completion options are displayed when the skill supports them; help exits without mutation.

Troubleshooting

This skill is the repo-owned workflow for the official Galileo MCP Server. It is render-first and client-tooling focused: it prepares MCP client configuration, validates that no secrets were written, inventories the live MCP tool surface, and points broader Galileo/Splunk work to the existing skills.

Required Intake

Before rendering, validating, doctoring, probing, or auditing this skill, ask the user for the Galileo instance console URL and record the exact value they provide, for example https://console.demo-v2.galileocloud.io/. Do not assume https://api.galileo.ai/mcp/http/mcp unless the user explicitly confirms the default Galileo Cloud instance.

Pass the URL as --galileo-console-url "$GALILEO_CONSOLE_URL" or set galileo.console_url in the spec. The renderer derives the MCP URL from that console URL unless the user provides a specific --mcp-url. HTTPS is required for every non-loopback console and MCP endpoint; cleartext HTTP is accepted only for localhost, *.localhost, or loopback IP validation fixtures.

Supported Paths

1. Client setup: render Cursor, VS Code, Codex, Claude Code, and AWS Kiro configs for https://api.galileo.ai/mcp/http/mcp or a self-hosted Galileo deployment.
2. Secret-safe bridges: use environment placeholders or owner-only local .env.galileo-mcp bridge files; never inline Galileo API keys. The rendered dependency-free Node.js bridge translates stdio JSON-RPC to Streamable HTTP directly; it does not rely on mcp-remote.
3. Tool inventory: probe initialize, tools/list, prompts/list, and resources/list without credentials, then compare live server identity, version, tool names, and schemas with the checked-in catalog.
4. Risk gating: treat dataset and prompt creation tools as write/generation tools that require explicit operator review.
5. Observability handoff: render a Python MCP client + add_tool_span handoff for applications that call MCP servers and need Galileo tool-span logging.
6. Product boundaries:
   • Full Galileo lifecycle and Splunk wiring: galileo-platform-setup
   • Agent Control and Cursor hook governance: galileo-agent-control-setup
   • Splunk HEC/OTLP, dashboards, and detectors: existing Splunk skills
   • Dataset versioning/collaboration, experiment groups/ranking, metric recomputation, SQL/Text-to-SQL metrics, Agent Graph analytics, saved views, Protect, Luna/Luna Studio, Trends, annotations, feedback, Python/TypeScript SDK reference work, provider/cost management, and any other Galileo capability outside the live MCP tool catalog: explicit handoff, not silent omission

July 7-August 7, 2026 product boundaries

The reviewed July 7, July 21, August 4, and August 7 releases add platform and documentation boundaries, not new tools in the observed 9-tool MCP catalog. Keep these explicit when planning MCP client setup:

• AI Assistant is an enterprise beta console feature that requires a configured LLM integration and support enablement. It is currently read-only and has no documented public Assistant API or MCP tool. It now works throughout Galileo debugging and prioritizes signals by criticality. Hand off readiness, enablement, and console evidence to galileo-platform-setup.
• Global dashboards span projects and log streams in the console. The documented public Trends API remains project/log-stream scoped, so do not claim global dashboard CRUD automation; hand off UI readiness and evidence.
• Generic alert webhooks support None, Bearer, or Basic authentication and a version 1.0 payload, but no public webhook/alert CRUD API or MCP tool is documented. Hand off receiver or relay design, configuration, and validation.
• Experiment groups require Galileo Python SDK 2.2.0 or later. The MCP setup_galileo_experiment tool provides guidance only; group lifecycle, comparison, and ranking remain a platform handoff.
• Large-dataset Playground and experiment metric processing now uses batching. Galileo does not document an exact maximum or client-side batch-size control; MCP dataset creation/status does not automate batched experiment execution.
• Annotation Queues are generally available, but the observed MCP catalog has no queue lifecycle tool. Hand off queue access, assignment, annotation, export, and validation to galileo-platform-setup.
• AI-assisted custom-code metrics, organization billing usage, Model Pricing, Integration Costs, Trace Count alerts, and multimodal evaluation metrics are platform or console workflows outside this MCP setup skill.
• GPT 5.6 Sol, Terra, and Luna availability and console theme selection are platform state or operator preference, not MCP configuration capabilities.
• As of August 7, Galileo is named Splunk Agent Observability. The legacy docs apply to customers onboarded before that date; customers onboarded later use agent-observability-docs.splunk.com. For an August 7 onboarding, verify the tenant's linked documentation rather than inferring the documentation epoch.

Safe First Command

bash skills/galileo-mcp-server-setup/scripts/setup.sh --help

Primary Workflow

Render the full client matrix:

bash skills/galileo-mcp-server-setup/scripts/setup.sh \
  --render \
  --galileo-console-url "$GALILEO_CONSOLE_URL" \
  --client cursor,claude,codex,vscode,kiro \
  --output-dir galileo-mcp-rendered

Render from the non-secret intake template:

bash skills/galileo-mcp-server-setup/scripts/setup.sh \
  --render \
  --spec skills/galileo-mcp-server-setup/template.example \
  --output-dir galileo-mcp-rendered

Render from a self-hosted Galileo console URL:

bash skills/galileo-mcp-server-setup/scripts/setup.sh \
  --render \
  --galileo-console-url https://console.galileo.example.com \
  --output-dir galileo-mcp-rendered

Validate rendered files:

bash skills/galileo-mcp-server-setup/scripts/validate.sh \
  --output-dir galileo-mcp-rendered

Probe live MCP metadata without credentials:

python3 skills/galileo-mcp-server-setup/scripts/probe_mcp.py \
  --mcp-url https://api.galileo.ai/mcp/http/mcp

Optionally verify an API key with a read-only /v2/current_user check:

chmod 600 /tmp/galileo_api_key
python3 skills/galileo-mcp-server-setup/scripts/probe_mcp.py \
  --auth-check \
  --galileo-api-key-file /tmp/galileo_api_key

CLI Contract

setup.sh supports --render, --validate, --doctor, --probe, --apply, --dry-run, --json, --client, --spec, --output-dir, --mcp-url, --galileo-console-url, --galileo-api-key-file, --accept-galileo-mcp-write-tools, and --allow-loose-key-perms. Client aliases include all, claude-code, vs-code, and aws-kiro.

--apply --client codex writes a local file-backed bridge environment and runs the rendered Codex registration helper. Cursor, VS Code, Claude, and Kiro config merges remain explicit reviewed handoffs and fail closed under automated apply.

Secret Handling

Use file-based or client-side secret injection only:

• --galileo-api-key-file for optional validation/probe checks and Codex apply
• ${env:GALILEO_API_KEY} for Cursor and local bridge workflows
• ${input:galileo-api-key} for VS Code prompt-string workflows
• .env.galileo-mcp local-only bridge file for Codex, Claude Code, and Kiro

Run chmod 600 mcp/.env.galileo-mcp after copying the example. The bridge also checks a referenced GALILEO_API_KEY_FILE for owner-only permissions. The lab-only GALILEO_MCP_ALLOW_LOOSE_KEY_PERMS=1 override is never rendered or enabled automatically.

The bridge sends Accept: application/json, text/event-stream, captures and propagates Mcp-Session-Id, sends the negotiated MCP-Protocol-Version, and supports JSON and SSE responses for initialize, notifications, tools, prompts, and resources. HTTP redirects are rejected so the Galileo API key cannot be forwarded to another origin. The initialize/initialized exchange is ordered; subsequent POSTs are concurrent so notifications/cancelled can bypass the tool request it cancels. Server SSE is bounded per event and reconnects with capped exponential backoff; a server that rejects GET disables that optional stream without affecting POST calls.

Never pass API keys in chat or argv. Direct secret flags such as --galileo-api-key, --api-key, --token, --password, and --authorization are rejected.

The optional authenticated /v2/current_user probe rejects HTTP redirects and returns only status plus response-key names. It never forwards the Galileo-API-Key header to a redirect target or includes its value in errors.

Tool Groups

• Guidance/public: search_docs, integrate_galileo_with_openai, integrate_galileo_with_langchain, setup_galileo_experiment
• Tenant read: get_logstream_insights, get_logstream_signals, validate_dataset
• Tenant write/generation: create_galileo_dataset, create_prompt_template

Unknown future tools are treated as manual-approval-only until the catalog is updated.

Validation

Run the complete audit gate:

bash skills/galileo-mcp-server-setup/scripts/deep_audit.sh

For deterministic local-only validation without live Galileo network checks:

bash skills/galileo-mcp-server-setup/scripts/deep_audit.sh --skip-live

Focused rendered-output validation:

bash skills/galileo-mcp-server-setup/scripts/validate.sh \
  --output-dir galileo-mcp-rendered
python3 -m py_compile \
  skills/galileo-mcp-server-setup/scripts/render_assets.py \
  skills/galileo-mcp-server-setup/scripts/probe_mcp.py \
  skills/galileo-mcp-server-setup/scripts/audit_product_coverage.py

See reference.md and the files under references/ for the client matrix, tool catalog, product gap matrix, and troubleshooting notes.