What Claude and ChatGPT can do with CodeHerder
What a claude.ai or ChatGPT connected app to CodeHerder can read and write on your behalf, what it can't touch, and what to do if something goes wrong.
A connected app is a live link from claude.ai or chatgpt.com to CodeHerder, using your own account and your own permissions — connecting one never gives Claude or ChatGPT any access you don’t already have. If you can’t see a workspace, a task, or a page in the workspace wiki in the CodeHerder app, the app can’t see it either.
This page describes what each capability actually does. For the click-by-click steps to add the app itself, see Connect CodeHerder to claude.ai or ChatGPT.
What it can do
Read
See your workspaces. Lists the CodeHerder workspaces you’re a member of (up to 200, most recently created first), or checks whether you have access to one specific workspace. If a workspace’s account requires SSO and you haven’t signed in through it recently, the list still shows that workspace’s name and place in the hierarchy, marked as needing SSO. It leaves out the rest. See When your account requires SSO.
Say something like: “What CodeHerder workspaces do I have access to?”
List tasks. Shows tasks in a workspace, including its sub-workspaces. By default this only shows open work (not-yet-started and in-progress tasks) — ask for completed tasks explicitly if you want those too. You can also filter by task type, such as bugs.
Say something like: “What are my open tasks in the Marketing workspace?” or “Show me the bugs that are done in Backend.”
Read one task in full. Pulls up everything about a single task: its title, description, status, priority, its fields (such as acceptance criteria), and the results of any checks recorded on it. Each acceptance item shows its proof stage and whether it is due, deferred or ticked.
Say something like: “Give me the full details on task abc-123.”
See a task’s sub-tasks. Lists the tasks filed directly under one task, oldest first. It shows sub-tasks in every status unless you ask for open ones only. A sub-task in a workspace you can’t see is left out. See Task hierarchy.
Say something like: “What sub-tasks does task abc-123 have?”
See what an agent asked on a task. Lists the questions and heads-up notes agents filed on a task, oldest first. By default you get the open ones. Ask for answered ones too if you want the history. Each open question says whether it is holding the task up. See Agent input.
Say something like: “Is any agent waiting on an answer on task abc-123?”
See why a task is blocked. Lists the blockers on a task, newest first. By default you get the unresolved ones, including a task that is waiting on another task. Ask for resolved ones too if you want the history. See Blockers.
Say something like: “Why is task abc-123 blocked?”
Read a task’s comments. Shows the discussion thread on a task, newest first. A comment mirrored from a linked pull request or merge request shows where it came from. See Pull request and merge request discussion.
Say something like: “What’s the latest discussion on that task?”
Read recent activity. Shows a feed of what’s happened recently — your own activity, everything in a workspace, or everything on one specific task. Your own activity follows the same SSO rule as everything else (see below).
Say something like: “What’s happened on this task recently?” or “What have I been up to in CodeHerder today?”
Read the team’s workspace wiki. Your team can save durable notes, conventions, and context in CodeHerder’s workspace wiki so agents and teammates don’t have to relearn the same things. Claude/ChatGPT can list what’s there and read one page in full. You see pages shared in the workspace and in its parent groups — not another teammate’s private notes.
Say something like: “What does our team’s CodeHerder wiki say about deploy conventions?”
Search everything you can see. Runs the same global search the CodeHerder app’s own search box does — across workspaces, tasks, the workspace wiki, messages, task comments, and the built-in help docs — and comes back with matches grouped by what kind of thing they are. Only ever searches what you already have access to.
Say something like: “Search CodeHerder for anything about the login bug.”
Read the CodeHerder guides. Lists or reads the built-in support/help docs — the same ones you’d find in the CodeHerder app’s own help section — so Claude/ChatGPT can answer “how do I…” questions about CodeHerder itself using the real, current docs instead of guessing.
Say something like: “What does the CodeHerder guide say about connecting a device?”
Write
Create a task. Creates a new task in a workspace, following that workspace’s own task rules
(what types it allows, what priorities are valid, and so on). You can also set its type, priority,
a parent task to file it under, field values such as acceptance criteria, and other tasks it
depends on, all in the same request. An acceptance item may name a proof stage. MCP has no per-item write: use the app or ch task acceptance to change one item later.
Say something like: “Create a high-priority bug in the Backend workspace called ‘Fix the flaky login test’, with acceptance criteria: the test passes ten runs in a row in CI.”
Comment on a task. Posts a comment on an existing task — the same as posting one in the app. Anyone mentioned in the comment gets notified, people watching the task get notified, and if an agent is actively working on that task, it gets woken up to see the new comment.
Say something like: “Add a comment to task abc-123 saying the fix is ready for review.”
Retrying is safe. Asking for the exact same new task or comment twice within a couple of minutes, for instance after a request seems to hang, creates it only once.
Link one task to another. Makes an existing task depend on another existing task, the same as adding a dependency in the app — the depending task isn’t ready to start until the one it depends on is done. Asking to link two tasks that are already linked doesn’t create a duplicate. Linking a task to itself, or in a way that would create a loop, is refused, and so is a task you don’t have access to.
Say something like: “Make task abc-123 depend on task def-456.”
What it cannot do
The app is deliberately limited to reading and to the three actions above. It cannot:
- Start, stop, or otherwise drive an agent session or sandbox
- Change a task’s status or move it through its workflow
- Edit a workspace’s workflows
- Edit or delete tasks or comments
- Answer an agent’s question or acknowledge its note
- Resolve or file a blocker
- Write wiki pages or send messages
- See another person’s or agent’s private wiki page
- See any workspace you are not a member of
Read-only vs read-and-write access
When you connect, CodeHerder shows you an Allow access to your account? screen that lists exactly what the app will be able to do — see What sign-in looks like for what that screen shows. A normal claude.ai or chatgpt.com connection asks for both read and write, so approving it gives you all three write actions above right away, with no extra prompt.
To check which access a connection has, open Connected apps on your My work page. The Access column shows Read only or Read and write. On a read-only connection, every write action is refused. Reconnecting won’t change that. Remove the app and add it again, and approve the access it asks for.
When your account requires SSO
If your account turns on Require SSO, a connected app acts as you, so it needs your recent SSO sign-in like everything else you do in that account. Without one:
- A request about that account’s workspaces or tasks fails with an SSO required error. The error says “Sign in with your account SSO to continue” and gives a sign-in link. Open the link in your browser, sign in through SSO, and ask again. You don’t need to reconnect the app.
- Your workspace list shows the account’s workspaces marked as needing SSO.
- Asking for your own activity fails the same way.
Your CodeHerder administrator sets how long an SSO sign-in counts as recent. When it runs out, sign in through SSO again.
If your account also turns on SSO-only sign-in, you can’t connect the app by signing in another way. See Connect CodeHerder to claude.ai or ChatGPT.
Limits you may notice
- Long lists come back a page at a time. If a workspace has a lot of tasks, or a task has a lot of sub-tasks, comments, questions, or blockers, you’ll get the first batch — just ask for the next page (e.g. “show me more”) and Claude/ChatGPT will fetch it.
- The team wiki has a cap here. When Claude/ChatGPT lists wiki pages through the app, only
a bounded number are recalled at once; if your team has saved a lot, you may be told the list was
cut off. (This cap is specific to the app — the
chCLI’s own page listing has no such limit.) - A burst of requests can be rate-limited briefly. If you (or an automation) fire off a lot of CodeHerder requests in a short window, you may see a brief “too many requests” message — it clears on its own within a minute.
If something goes wrong
- You get “not found” for a task or workspace you can clearly see in the CodeHerder app. You’re probably signed into a different CodeHerder account, or the item belongs to a workspace you’re not a member of under that account. Check which account the app is signed in as.
- You get an “SSO required” error. Your account requires SSO and your last SSO sign-in is too old. Open the sign-in link in the error, sign in, then retry. See When your account requires SSO.
- A create-task, comment, or link request is refused. The connection is probably read-only. Check the Access column under Connected apps (see above).
- A sub-task you expect is missing from a task’s sub-task list. It probably sits in a workspace you aren’t a member of. The app leaves those out.
- Tools stop being offered partway through a ChatGPT conversation, even though the app is still listed as connected. This is a known ChatGPT-side issue, not a CodeHerder problem — start a new chat and it should work again.
If you’re asked to sign in every time, or setup fails, see the troubleshooting list in Connect CodeHerder to claude.ai or ChatGPT. To see every app you’ve approved and revoke one, see My work.
Last updated