Drive your herd from claude.ai and chatgpt.com
You can now drive your CodeHerder herd from claude.ai and chatgpt.com: the places where a CLI can’t reach.
That’s a more literal statement than it sounds. A web app has no shell, no PATH, no local
process it can spawn. ch is excellent everywhere it can run, on your laptop, in CI, on a bare
device, but it needs somewhere to run. Before this, that meant the entire product was
unreachable from two places a lot of people now spend most of their working day: a browser tab
with claude.ai open, or one with chatgpt.com open. If your herd’s status lived only behind a
terminal, it didn’t live in the conversation you were already having.
That gap is closed. CodeHerder is now reachable as a remote MCP connector, and once it’s added to claude.ai or chatgpt.com, you can ask either one what your herd did today, triage the backlog out loud, file a task from a conversation you’re already in, or post a comment on one that’s stuck, all without opening a terminal or the CodeHerder app itself.
What this actually unlocks
Picture the parts of a workday that already happen in a chat window. You’re talking through a bug with an assistant, and the fix belongs on the backlog, not in your head five minutes later. You want to know if last night’s overnight run finished before you’ve even opened your laptop. Someone asks what a task is blocked on and the honest answer is “let me check,” except now you don’t have to switch apps to check. None of that needs a new interface. It needs the interface you’re already in to be able to reach the herd.
What ships in v1, stated plainly
MCP is a widely adopted protocol, and plenty of tools already speak it. The interesting part is what CodeHerder puts behind the protocol: a governed, multi-tenant herd with per-task cost attribution and stage gates, not a bare list of API calls with an MCP wrapper taped on.
v1 is deliberately narrow, and the boundary is worth stating rather than leaving implied. You can read: list your tasks and workspaces, read activity on a task or a workspace, pull up a workspace memory entry. You can write in exactly two ways: create a task, and post a comment on one. That’s it. You cannot spawn an agent session, drive a sandbox, or edit a task-type schema through this connector, not yet. Seven of the nine tools are read-only; the two that write both require a credential scoped for it, and a read-only connection is refused outright if it tries either one.
Setting up the connector
Both walkthroughs below assume nothing: you don’t need to have used CodeHerder’s own app before, and you don’t need a separate password for either client. Signing in through the connector is your normal CodeHerder login, just triggered from inside claude.ai or chatgpt.com instead of from a browser tab pointed at CodeHerder directly.
Connect claude.ai
-
In claude.ai, open Settings, then Connectors.
-
Click Add custom connector.
-
Paste in this URL exactly, no trailing slash and no
www.:https://api.codeherder.com/v1/mcp -
A sign-in popup opens. This is your normal CodeHerder identity provider’s own page, the same one you’d see logging in any other way. Nothing here asks for a CodeHerder password directly; you authenticate through the popup, and Claude never sees your credentials.
-
Once the popup closes, the connector shows as connected. Start a new chat and ask something concrete, like “what are my open CodeHerder tasks?” A real answer, with real task titles, is how you know it worked.
Connect chatgpt.com
-
In chatgpt.com, open Settings, then Connectors, then Advanced, and turn on Developer mode. Custom connectors are gated behind this setting; without it, there’s no option to add one.
-
Back in Connectors, click Add connector (some versions of the UI label this Create).
-
Paste in the same URL:
https://api.codeherder.com/v1/mcp -
Sign in the same way as above: a popup carrying the CodeHerder identity provider’s own page, not a ChatGPT-branded form.
-
Confirm it worked with a real question, for example “read the latest activity in my CodeHerder workspace.”
When something doesn’t work
A connector added but showing no response, or an error where a normal answer should be, usually traces to one of a small number of causes.
The connector shows as added, but nothing it asks ever comes back authorized. This almost always means the sign-in step was skipped or dismissed partway through: remove the connector and add it again, and this time complete the popup all the way to the confirmation screen before switching back to the chat.
Sign-in opens, but instead of returning you to the chat it lands on an error page. That’s a sign the connector’s registration went stale, most often because it sat unused for a while. Removing it and adding it fresh re-registers it cleanly; there’s no state to preserve, since nothing about your CodeHerder account changes when you do.
You’re asked to sign in again every single session instead of staying connected. That’s an expired credential, not a broken one: reconnecting issues a fresh one and the prompt stops.
The connector answers, but only for a workspace you don’t recognize, or refuses to see a task you know exists. The connector only ever sees what your CodeHerder login already sees. If a workspace is missing, that’s a membership question for whoever administers it, not a connector bug.
One more, specific to chatgpt.com: a connector that works for the first question in a chat and then quietly stops being offered to the model at all, even though it’s still listed as connected. We chased this down during our own acceptance testing and confirmed it’s a ChatGPT-platform issue, not a CodeHerder one; it reproduces against other vendors’ connectors too, not just ours. If you hit it, start a new chat. There’s no setting on the CodeHerder side that changes it.
The first useful call, and where to go deeper
The fastest way to know a connector is real: ask for your open tasks, or ask it to post a comment on one you already have open. From there, the click-by-click setup guide is here: Connect CodeHerder to claude.ai or ChatGPT, kept current as more tools are added, so this post doesn’t have to be.
