Postlint MCP Server

An MCP server that answers one question: does this post fit?

A language model cannot count characters by inspection, and on these platforms neither can you. The limits are not what they look like. X bills every URL at 23 characters through t.co whether the link is 12 characters or 200. Bluesky counts extended grapheme clusters, so a four-person family emoji is 1 and not 11. Mastodon charges nothing for the domain on a remote mention. Getting any of that wrong shows up as a rejected post, or a truncated one, at publish time.

Counting is what a tool call is for. The model cannot do it by inspection, and a deterministic function can do it exactly.

Why this exists. Two posts went out of a podcast promo workflow over the limit. A Bluesky post shipped at 302 against 300, with the line "Under 300 graphemes. Audit clean." sitting directly beneath it. An X post was drafted at 308 against 280 and would have been rejected on launch morning. Both were invisible to eyeballing, because in both cases the count was a claim and not a measurement. Both are regression tests in this repo.

Tools

Responses are small on purpose. check_post_all omits the arithmetic on passing rows because agents pay tokens per response.

How each platform counts

Every number above traces to a published source. Widely repeated figures that no primary source states — the Facebook post limit, the YouTube community post limit, Reddit's title cap, Instagram's organic caption cap — are deliberately absent. A limit that cannot be defended makes a passing check worth nothing.

Setup

Published on npm. The config blocks below use npx, which fetches it on first run; no clone required.

git clone https://github.com/conorbronsdon/postlint-mcp.gitcd postlint-mcpnpm installnpm run build

Claude Code

Add to your .mcp.json:

{  "mcpServers": {    "postlint": {      "command": "node",      "args": ["/absolute/path/to/postlint-mcp/dist/index.js"]    }  }}

Claude Desktop

Same block, in claude_desktop_config.json.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.postlint]command = "npx"args = ["-y", "@conorbronsdon/postlint-mcp"]

No token, no environment variables, no network access. Once the package is published, npx -y @conorbronsdon/postlint-mcp replaces the node invocation everywhere above.

Verify

Ask your assistant: "Check this post for X and Bluesky," and paste something with a couple of links in it.

A worked example

The X post that started this, run through check_post with platform: "x":

{  "platform": "x",  "limit": 280,  "unit": "weighted characters",  "length": 308,  "over": true,  "remaining": -28,  "drivers": [    "3 URLs counted as 23 each = 69",    "239 other characters counted as 1 each"  ],  "warnings": []}

The drivers line is the useful part. 69 of the budget went to links before a word was written, which tells you to move two of them into a reply rather than trimming prose.

The same post through check_post_all:

{  "fits": ["x_premium", "linkedin", "threads", "mastodon", "discord"],  "over": ["x", "bluesky"],  "rows": [    { "platform": "x", "length": 308, "limit": 280, "over": true, "drivers": ["3 URLs counted as 23 each = 69", "239 other characters counted as 1 each"] },    { "platform": "bluesky", "length": 330, "limit": 300, "over": true, "drivers": ["3 URLs counted in full = 91 (Bluesky does not shorten links)", "239 other graphemes"] },    { "platform": "mastodon", "length": 308, "limit": 500, "over": false, "remaining": 192 }  ]}

One post, three different lengths — 308, 330, and 308 again — from the same 330 characters of text. That gap is the whole reason this exists.

Draft placeholders

Drafts carry link placeholders, and [URL] is five characters while a real link is not. A post measured with the placeholder in place and posted with the link filled in is a post measured wrong; one draft came in at 264 that way and posted at 282.

So [URL], [LINK], [YOUTUBE URL], [SUBSTACK URL], and similar are priced as a real link (a 28-character YouTube short link, the shortest thing normally posted) and the response carries a warning saying the count is a floor.

What it does not do

• It does not post anything. There is no write path, no credential, and no network call of any kind. That last one is enforced rather than asserted: a test replaces fetch, XMLHttpRequest, and WebSocket with throws and drives every tool, so a call added later fails CI instead of quietly making this sentence false.
• It does not check an instance's actual limit. Mastodon servers configure their own; this reports the 500 default and tells you to read configuration.statuses.max_characters from the target server yourself.
• It does not truncate. A truncate_to helper was considered and left out. Cutting a post at a character offset splits URLs, breaks grapheme clusters, and lands mid-sentence, and cutting it at a "safe" boundary silently drops whichever clause happened to be last. Either way the tool would be deciding what the post says. It reports the number and leaves the edit to you.
• It does not detect every URL a platform would. Links with a scheme and www.-prefixed hosts always match. A bare domain matches only on a common TLD (src/count.ts holds the list), where the real twitter-text implementation carries the full IANA registry. Write https:// in front of a link and the count is exact.
• It does not count media, polls, quote posts, or link cards. Those have their own rules and this measures text.
• It does not know about content warnings. On Mastodon a CW counts toward the same 500. This checks the body alone.
• It does not carry limits it cannot source. See the platform table.

Development

npm installnpm run buildnpm test

Tests make no network calls, because the server makes none. The two historical over-limit posts are regression fixtures in src/__tests__/lint.test.ts, alongside grapheme cases for ZWJ family emoji, regional-indicator flags, skin-tone modifiers, combining accents, and CJK.

Contributing

Issues and pull requests are welcome. A new platform needs three things: the limit, the unit it is measured in, and a published source. A new counting rule needs a test that fails without it. Numbers repeated by third parties are not sources.

About

Built and maintained by Conor Bronsdon. I host the Chain of Thought podcast, which covers AI infrastructure, developer tools, and how practitioners actually use this stuff. I built this after shipping two over-limit posts in a workflow that was supposed to catch them.

Companion tools:

• op3-mcp: podcast analytics through OP3 — downloads, geography, apps, per-episode breakdowns.
• podcastindex-mcp: the Podcast Index MCP server, search by person or topic, trending shows, feed health.
• substack-mcp: read posts and manage drafts on Substack, safe for agent workflows.
• Transistor MCP: Transistor.fm's official MCP server. Episodes, publishing, and analytics.
• ai-tools-for-creators: a curated list of AI skills and MCP servers for people who ship ideas for a living.

More at chainofthought.show and on X.

Disclaimer

This is an independent personal project, not affiliated with, sponsored by, or endorsed by any company. All views expressed are my own.

License

Apache-2.0