Follow Discord
Sweep 19 Sep 2026 · 02:36Z Build v2.1.278 500 read Stable v2.1.267 Latest v2.1.278 Next v2.1.278 Feeds RSS JSON llms.txt Unofficial
Home
MCP server

The whole corpus, in your client.

One URL, added once, and your assistant can read every Claude Code release this site has taken apart, every flag and setting name mined out of the binary, Anthropic's own documentation, and what each build actually told the model its tools do.

No API key, no account, no rate limit worth thinking about. Everything it answers is a page you could open yourself.

Tools15the same 15 the plugin carries
API keys0nothing here is behind a login
Accounts0it never learns who you are
TransportHTTPJSON-RPC over POST, no stream
The addresshttps://changelogs.core-directive.com/mcpthat is the whole install

Add it in one line

no key, no account, no config file to create

Claude Code

One command

claude mcp add --transport http cc-changelog https://changelogs.core-directive.com/mcp

Add --scope user to have it in every project rather than this one.

Claude Desktop

Settings → Connectors → Add custom connector

https://changelogs.core-directive.com/mcp

Paste the URL. There is nothing to authorise: every tool here is a public read.

Cursor

In ~/.cursor/mcp.json

{"mcpServers":{"cc-changelog":{"url":"https://changelogs.core-directive.com/mcp"}}}

Project-local instead: .cursor/mcp.json in the repository.

VS Code

One command

code --add-mcp '{"name":"cc-changelog","type":"http","url":"https://changelogs.core-directive.com/mcp"}'

Or add the same object under mcp.servers in your settings.

Anything else

Streamable HTTP

https://changelogs.core-directive.com/mcp

One URL, no key, no account. It speaks MCP over POST and opens no event stream.

What it can tell you

15 tools, one corpus each
ToolAnswersWhat it is for
changelog One release, as a document The whole changelog for a version, read straight out of that build rather than from a release note.
search Every release at once The fastest way in when you know a word but not a version: a flag, a subsystem, a command.
releases What versions exist Dates and counts, newest first, so a model can find the release it should be reading.
upgrade What changed since your build Name the version you were on and the one you are on; everything in between comes back at once.
entries One release, as findings The list rather than the prose, filterable by section, tier and area.
entry One finding, in full The whole of what was written about a single change, with its evidence.
reference What a name means Every environment variable, flag, settings key, slash command, tool and hook mined from the binary.
docs Anthropic's own docs The published documentation as this site captured it, searchable and quotable by page.
docschanges What Anthropic quietly edited The same docs read on a timer and diffed, so a page that gained a paragraph last Tuesday says so.
blog The posts on this site The one part of the corpus that is opinion, and it is labelled as such.
prompts What the model was told The stock tool descriptions a build shipped, release by release, as the binary carries them.
modsapi The plugin runtime's API The undocumented hooks API, mined from the newest build: what is on `$`, and what an event carries.
mods How to write a hooks module The ten-page guide to the runtime nothing upstream documents: tiers, events, panes, gotchas.
stories A feature, across releases Five releases that each touched one thing, grouped by hand into the order they actually arrived.
watch Which entries have aged Flags that moved, docs that arrived, corrections made since. Added and dated, never swapped in.

Each of these stands in front of a JSON route this site already publishes, and every answer carries the source_url of the page that produced it, so an assistant can hand you the address rather than just the text.

What it cannot do

it reads, and that is the whole of it
The claimWhy it holds
NoIt cannot write anything.Every tool resolves to a published GET route. There is no verb here that changes a row, and no code path that could.
NoIt cannot read your project.It is a remote server on a URL. It receives the arguments your client sends and has no other view of your machine.
NoIt does not know who you are.There is no key to issue and no account to make, so there is nothing to attach a call to. Requests are counted per address and nothing else.
NoIt is not Anthropic's.This site is unofficial and unaffiliated. What it reports about a release is read out of the shipped build; where Anthropic wrote something too, it is quoted and labelled as theirs.
YesEverything it answers is already public.Each tool is a JSON route you can open in a browser. The server exists so a model does not have to be told the shape of the site first.

How it works

MCP 2025-06-18, stateless
FactWhat it means
Streamable HTTPOne POST per call, carrying a JSON-RPC message. A GET asking for an event stream is answered 405, because there is no stream to open.
StatelessNo session id and nothing held between calls, so a reconnect costs nothing and a dropped connection loses nothing.
Tools onlyNo resources, no prompts, no sampling, no elicitation. The capability object says so rather than leaving your client to find out.
A wrong guess is an answerA name that does not exist comes back as a readable failure naming the shape it should have had, not as a broken call, so a model can correct itself in one retry.
Same rate limit as the APIIt shares the read limit every JSON route here carries. Normal use does not come near it.

If you use Claude Code

there is a second way in

The cc-changelog plugin carries the same 15 tools and adds a release pane, so a new version announces itself while you work. The MCP server is the same corpus with the install step removed: one address, any client. Running both is harmless, but your model will see each tool twice.

This server is listed on Smithery, if that is where you look for servers. The address above is the same one either way.