Engineering MCP: Why Your Agent Can't Find Tools and How to Fix It
k_man9 · reddit · 2026-08-04
The author shares troubleshooting experiences and solutions when integrating custom MCP servers with AI agents like Lovable.
The Problem
Despite having well-configured MCP tools and Markdown Playbooks, Lovable repeatedly failed (e.g., guessing API hostnames, triggering CORS rules). The root causes were that Lovable doesn't support MCP Resources, and its client failed to properly rank instructions (truncating alphabetically).
The Solution
To ensure agents reliably discover and use documentation, the author proposes a 3-layer redundant discovery mechanism:
- Instructions on connect: Ordered by importance, pinning critical playbooks first.
- Dual-transport index: Exposing both resources/list and synthetic skillslist tools to support clients lacking MCP Resources support.
- In-tool pointers: Embedding documentation URIs directly inside individual tool descriptions (e.g., see skill://...).
When dealing with 50+ MCP tools, this intentional redundancy is crucial for stable agent performance.
More from coding & agent
- Scale AI Proposes Interaction-Centric Taxonomy to Localize Agent Failures — ScaleAI · 2026-08-04
- Developers Shift from Finding the 'Best' AI Coder to Task-Specific Models — ingliguori · 2026-08-04
- AI Agent Skill Automates 27 Open-Source Docs Generation with Real Context — zyphraxns · 2026-08-04
- Time Management in the Agent Era: Maintaining Human-Machine Asymmetry — TheZachMueller · 2026-08-04
- browser-use releases video-use: editing videos via coding agents — browser-use · 2026-08-04
- Compound Engineering plugin supports Claude Code and Cursor — EveryInc · 2026-08-04