Agents read less, act, and get content from anywhere

mira mcp keeps one build for the whole session and rebuilds a project only when one of its files changes. Calls after the first answer in about a millisecond, down from 120 to 150 ms.
The tools are renamed and described in fewer words: site, pages, search, read, items, data, and media.
search returns five matches by default, each pointing at the section that matches, with a snippet of at most 160 characters.
read returns a whole page or one section, as path#section, and a long page is cut with a list of its remaining sections.
items filters, sorts, and picks fields from a collection before anything is sent. Collection schemas take time, datetime, object, number[], and object[], so hours, schedules, and prices can be queried.
llms.txt lists each page’s size in tokens, so agents fetch only what they need.
mira audit --agent reports what each page costs an agent to read and flags pages over a budget.
Markdown copies of pages leave out footers, forms, hidden elements, and decoration.

On this site, listing the changelog’s releases with three fields went from 334 tokens to 63, and a search from 359 to 246.

Declare actions, such as book_table, with typed input and an HTTPS endpoint. Each becomes an MCP tool and is listed in llms.txt.
Input is checked against its types, and nothing is sent until the person using the agent agrees, through their client’s own prompt or a one-time confirmation tied to the exact input.

Collections can come from Sanity, Contentful, Supabase, any GraphQL API, or any JSON endpoint, including functions on AWS, Google Cloud, and Azure. Rich text becomes Markdown, and its images are processed like local ones.
Pages and entries can be MDX, rendered at build time with your own .mira components. No JavaScript runs.

hosts.amplify writes the config for AWS Amplify Hosting.
Every Mira site can be read over MCP from its deployed URL with mira mcp --url.
mira dev and mira mcp no longer rewrite host config files in the project root.