MCP Servers
Adlyse exposes its performance-marketing data and judgments over the Model Context Protocol so external AI clients — Claude Code, Claude.ai, ChatGPT Desktop, or your own orchestration — can read from and write into Adlyse the same way our in-app assistant does.
Five MCP servers run behind a single combined endpoint. All are scoped to your organization by your API key. adlyse-admin is intentionally separate and high privilege; enable it only for trusted automation.
The servers#
| Server | Path | What it exposes |
|---|---|---|
adlyse-insights | /mcp/insights/ | ABI, waste, safe-budget, impact analysis, decision log, guidelines |
adlyse-reports | /mcp/reports/ | List / generate / fetch performance reports with presigned PDF URLs |
adlyse-ads | /mcp/ads/ | Cross-platform campaign & ad reads; gated writes (pause / budget) |
adlyse-workflows | /mcp/workflows/ | List and trigger your Adlyse workflows; poll run status |
adlyse-admin | /mcp/admin/ | Control-plane actions for assistants, handoffs, skills, and memory files |
- Production host:
https://mcp.adlyse.com - Local dev host:
http://localhost:8766(run./start-mcp.shfromserver/)
Getting started#
- Authentication — how API keys and OAuth work
- Quick start — wire Claude Code (or any MCP client) to Adlyse in under a minute
Conventions#
Every tool response carries a source tag — e.g. adlyse.dashboard_compute@v1 — so callers can verify which computation produced the data and reason about version drift.
Writes are preview-safe. Write tools (currently update_campaign_status and update_campaign_budget) accept a preview: true parameter that returns the before/after diff without calling the platform API. Apply with preview: false.
Writes auto-log. On a successful mutation, Adlyse writes a DecisionLogEntry with created_via="tool" and a short fingerprint of the calling API key. This feeds back into get_impact_analysis so future metric shifts correlate with past decisions. See Decision Log for the full model.
Admin is explicit. adlyse-admin is not meant as a default connector. It can create, edit, and delete configuration and memory, so OAuth tokens need the admin scope and users need workflow-management permission. Organization API keys continue to authenticate org-scoped admin access.
Errors are structured. Failed tool calls return { "error": "...", "error_type": "..." } rather than raising. Common error_type values: auth_error, not_found, validation_error, unsupported_for_platform, platform_error.