For agents.

This site is built to be read and used by AI agents as well as people. Everything an agent gets here says what the pages say. It is all read-only, and nothing sends a message on anyone’s behalf.

MCP server

Streamable HTTP at https://jasonlovell.ai/mcp. No sign-in. Read-only tools.

Claude Code

claude mcp add --transport http jasonlovell https://jasonlovell.ai/mcp

Claude and ChatGPT apps

Add it as a custom connector (remote MCP server URL).

https://jasonlovell.ai/mcp

Cursor, VS Code and other clients

{
  "mcpServers": {
    "jasonlovell": {
      "url": "https://jasonlovell.ai/mcp"
    }
  }
}

get_profile: Who Jason Lovell is: day role, career from mobile through Samsung, Captivate and PwC, how he builds, what he cares about now, and how to reach him.
list_studies: The ten research studies: each one's question, kind of result, a short outcome, date and page. Can filter by the question a study asks.
get_study: One study in full: the question, the result, the design decision, what was recorded, the evidence links and where the evidence stops.
list_builds: Public repositories by area, newest first.
get_build: One build: what it is and includes, and for builds with a design record, the question it answers, the design choice and where it stops.
list_releases: Specifications, a white paper and research software, with DOIs where they have them.
search_site: Find studies, builds and releases by keyword.
contact_routes: Ways to reach Jason, by purpose. It sends nothing: pass the route to the person you act for.

Signed tools

The MCP server’s tool definitions are published as a Tool Bill of Materials, in the TBOM specification I published in February. A client can check that the tools it connects to are the ones I signed, and that none has changed since.

Definitions: Each tool’s name, description, input schema and annotations, hashed with SHA-256 over RFC 8785 canonical JSON.
Signature: Ed25519, a detached JWS over the whole document, from the key published at /.well-known/tbom-keys.json.
Artifact: The server is hosted rather than installed, so its artifact is the tool surface itself: the hash of the tools array that tools/list returns.
Capabilities: Declared for every tool: no shell, file system, network or credential access, and no side effects.
Drift: The site’s tests check the live tools against the signed digests, so a changed tool fails them until it is signed again.

Check it here

In your browser: the signature with WebCrypto, then every tool against the live server.

Fetch the signed TBOM
Fetch the published signing key
Verify the Ed25519 signature
Compare every live tool with its signed digest
Compare the live tool surface with the signed artifact

Or check it yourself

With the specification’s reference implementation. It validates the document, then verifies the signature against the published key.

git clone https://github.com/jlov7/tbom-rfc && cd tbom-rfc
curl -sO https://jasonlovell.ai/tbom.json
curl -s -o tbom-keys.json https://jasonlovell.ai/.well-known/tbom-keys.json
uv run --no-project --with jsonschema --with cryptography --with jcs \
  python tbomctl.py check --schema tbom-schema-v1.0.2.json \
  --keys tbom-keys.json tbom.json

In the browser

Every page registers WebMCP tools for an agent working in the same tab (Chrome 146 and later). While an agent uses them, a small pill at the bottom of the page says what it is doing.

search_site: Search studies, builds, releases and pages by keyword.
read_page: Read any page of the site as Markdown.
open_page: Show a page to the person in the tab.
draft_message: Fill the contact form for the person to read and send. It never sends.

Plain text

Any page as Markdown: add .md to its path, or ask for it with an Accept header. Search engines index the HTML; the Markdown is for agents.

curl -H "Accept: text/markdown" https://jasonlovell.ai/work/realitybridge

llms.txtA map of the site for language models.
llms-full.txtEvery study, design record, build and release in one file.
index.mdThe home page as Markdown. Every page has one: add .md to its path.
feed.xmlStudies and releases as an Atom feed, newest first.
sitemap.xmlEvery page, with its last change.

Measured

Two coding agents answered 20 factual questions about this site, reading it three ways. Every run scored at least 19 of 20; Markdown and MCP took 3 to 81 times fewer input tokens than HTML. One run each, on 28 September 2026.

Correct answers, input tokens and time for each agent and way of reading the site
Read as	Claude CodeClaude Sonnet 5	CodexGPT-5.6 Sol
HTML pages	20/203.75M tokens289 s	20/20555k tokens106 s
Markdown twins	20/20105k tokens24 s	19/2090k tokens29 s
MCP server	20/2046k tokens14 s	20/20186k tokens36 s

Codex, reading the Markdown twins, answered “June 2026” to “In which month and year was the Tool Bill of Materials specification published?” An earlier run found a study count the HTML worded ambiguously; that wording is fixed. Input tokens include cached tokens. The questions, answer key and every answer.