Building a personal MCP server: what we learned
ORANO shipped a personal MCP server so subscribers can let their own AI agents read their saved context. These are the decisions that mattered, for anyone building something similar.
1. Read-only by default
We scoped the server to five read tools — list projects, get a project, get pending handoffs, search the library, read memory facts. No mutations. The blast radius of a buggy agent call is zero. Details: oranoai.com/mcp.
2. Personal keys, not shared tokens
Each user provisions personal keys (bearer auth, scope orano:read, up to 10 active). Nothing is shared across users, and revoking one key revokes one agent.
3. Budgets before features
A runaway agent loop is the most common failure mode. The budget is 240 calls per user per 60 minutes. Generous for real use, fatal for runaway loops.
4. Durable handoffs over polling
When a user explicitly sends a project to their agent, it lands in a durable per-key handoff queue rather than relying on the agent to poll at the right moment.
5. Publish a machine-readable manifest
oranoai.com/mcp.json documents the transport, auth format, onboarding URLs and tool list. Agents and integrators stop guessing.
6. Say what it can't do
No one-click OAuth yet, no write access, no cross-user reads. Honest limits are a feature: oranoai.com/llms.txt spells them out for both humans and models.
ORANO is on iOS and Android, $5/month.
Full guide on oranoai.com: How to connect your saved videos to Claude Code with MCP