Appearance
Integrations
Integrations connect agents to the systems your company already uses: Slack, Google Drive, Docs, Sheets, Calendar and Gmail, Notion, Jira and Confluence, GitHub, GitLab, Grafana, and anything else that speaks MCP.
They are set up per owner, in Settings → Integrations, by its owners and admins. The same list, and the same Add integration, also hold the apps that go the other way: people talking to agents from Slack and Telegram, marked Chat app. The tools every space has, web search, web fetch and Computer, are in the space's Settings instead.
Whose account the agent uses
This is the decision that matters, and it is made per integration:
| Mode | How it works | Use it when |
|---|---|---|
| Shared account | One token, set by an admin. Every agent acts as that account. | A service account with deliberately narrow access, e.g. a read-only Grafana viewer |
| Each person's token | Everyone pastes their own in Account → Connections | The system issues personal tokens and has no sign-in |
| Each person's account (sign-in) | Everyone clicks Connect, approves on the provider's page once, and the connection is kept | Slack, Google, Notion, Atlassian, GitHub, GitLab |
Connecting is one click
Integrations marked one click under Add integration need nothing filled in: choose one, sign in on the provider's page, and it comes back ready, connected, its tools loaded and on for your agents. Everyone else connects their own account from Account → Connections, and agents act as them.
- Notion, Jira & Confluence: CrewHall registers itself with the provider. On Notion you choose which pages to share; an Atlassian admin may have to allow CrewHall's domain first.
- GitHub and Google (Drive, Docs, Sheets, Calendar, Gmail): CrewHall uses the app people sign in with, when this server offers that sign-in. For Google, whoever hosts CrewHall enables each API on that app once (see hosting).
- The others open a short form: Slack needs a Slack app of yours (made in a click, below), GitLab the address of your GitLab, Grafana a token, and a custom MCP server its address. Most hosted MCP servers (Linear, Sentry, Asana…) then register CrewHall on the first Connect, followed by Refresh tools.
Each integration
What an agent can do is the tools you tick for it, described below. This is what adding each one asks of you.
Slack
Slack only lets apps made in your own workspace use its MCP server, so the Slack integration needs an app of yours. Create the Slack app, in the integration's form, opens Slack with everything filled in: the redirect URL, the permissions, PKCE and the MCP server. Pick your workspace, create it, Install App → Install to your workspace (a Slack admin may have to approve it), then copy its Client ID and Client Secret from Basic Information into the form.
- Just you: add Slack in your personal space, with your own app.
- An organization: an admin adds it once with the organization's app; everyone else just clicks Connect.
- When nobody can make an app for the whole workspace: the admin adds Slack with Each person's own app. Each person then creates their own the same way, and enters its client ID and secret when they click Connect. Disconnecting forgets it.
Agents see only what the person they act for can see in Slack. Talking to agents from Slack, the other direction, is the Slack chat app.
Google Drive, Docs, Sheets, Calendar and Gmail
Each product is its own integration, and each person connects their own Google account. Where this server offers Continue with Google, CrewHall uses that same app: whoever hosts it enables the product's API once (see hosting). Otherwise the form asks for an OAuth client from your Google Cloud project, with the redirect URL it shows. One client can serve every Google integration.
Drive can find, read and create files. Docs and Sheets read and edit their own files. Calendar reads events and free time. Gmail searches and reads mail, and can draft replies. Google may show an extra warning, or limit who can connect, until the app is verified for those scopes.
Notion
Nothing to create on Notion's side. Choosing Notion signs you in, CrewHall registers itself, and you come back connected. During that sign-in you choose which pages to share. Everyone else clicks Connect in Account → Connections and chooses their own pages.
Jira and Confluence
One integration covers both. CrewHall registers itself with Atlassian when you connect. A site or organization admin may have to allow CrewHall's domain first (Atlassian Administration, the Rovo MCP server). Each person then connects their own Atlassian account.
A shared service account is the other option: an admin pastes one API token, and every agent acts as that account. An Atlassian admin has to allow API-token authentication for the MCP server before that token works.
GitHub
Repositories, issues, pull requests, Actions runs and code search. Each person connects their own GitHub account, so an agent sees only what that person can see. Where this server offers Continue with GitHub, CrewHall uses that same app and there is nothing to create. Otherwise the form asks for a GitHub OAuth App (Settings → Developer settings), with the redirect URL it shows. A personal access token works too: Each person's token, or one Shared account token for everyone.
Private repositories is on unless you turn it off. On, a connection can read private repositories. Off, public ones only, and each person connects again for that to apply. The permission GitHub calls repo is also what allows writing, so write tools stay off until you tick them for an agent.
Granting an organization
Connecting approves the app for your personal account. It does not, by itself, approve it for an organization. Organizations turn on third-party access restrictions by default, and until someone grants the app, GitHub hides that organization's private repositories, issues and pull requests. Public repositories stay visible.
On the GitHub integration:
- Click Scan my GitHub account. CrewHall lists your personal account and each organization you belong to.
- Granted means this connection can use that organization's private data.
- Not granted yet means you belong to it, and GitHub is still hiding it.
- Click Open on GitHub. That opens the app under Settings → Applications → Authorized OAuth Apps. Under Organization access:
- Grant, if you own the organization. It takes effect immediately.
- Request, if you are a member. An owner approves it in the organization: Settings → Third-party Access → OAuth app policy. Until they do, a new scan still says Not granted yet.
- Scan again. The organization should now say Granted.
Organizations on the same card, left off, lets every agent use every organization this person's connection can see. Turn it on to tick the accounts agents in this space may use. On an agent, the gear then narrows that list, so one agent can be limited to one organization and another can keep all of them. A call that names anything else is refused.
Some organizations also require single sign-on for the app. The scan says so when GitHub asks for it; authorize it on the same GitHub page, then scan again.
A fine-grained personal access token includes only the organizations chosen when it was created (Settings → Developer settings → Personal access tokens). GitHub does not tell CrewHall which of those it was given, so the scan lists your memberships and leaves them unmarked. A classic token has its own Organization access list on the token's page, the same Grant or Request choice.
What the default GitHub tools include
The address CrewHall uses, https://api.githubcopilot.com/mcp/, includes repositories, issues, pull requests and users. Organization administration (members, teams, settings) is a separate GitHub toolset. To include it, set the integration's URL to https://api.githubcopilot.com/mcp/x/all, save, and Refresh tools. The organization still has to have granted the app, above, or those calls see nothing private.
GitLab
Issues, merge requests, pipelines and search, on GitLab.com or a GitLab you host. Change the host in the form for a self-managed instance. CrewHall registers itself when you connect.
GitLab only answers people in a top-level group that allows MCP. On GitLab.com, a group Owner opens the group, Settings → General → Permissions and group features, and turns on Allow connection to GitLab under MCP client access. Groups created recently have it on. On a self-managed GitLab, an administrator allows it for the whole instance (Admin → Settings → General → Visibility and access controls).
Grafana
Dashboards, Prometheus and Loki queries, alerts and incidents, through one shared service account. In Grafana, create a service account (Viewer is enough to read) and a token for it. Whoever hosts CrewHall sets GRAFANA_URL, GRAFANA_SERVICE_ACCOUNT_TOKEN and MCP_GRAFANA_SERVER_TOKEN, and runs docker compose --profile grafana up -d. Paste that MCP token into the integration as the shared token. Agents then act with the service account's permissions.
A custom MCP server
Any server that speaks MCP over HTTP: Linear, Sentry, Asana, or one you run. Enter its URL. If it supports the standard sign-in flow, CrewHall registers itself on the first Connect; otherwise enter a client ID and register the redirect URL the form shows. Servers that only run as a local process are not supported yet.
CrewHall's own apps
Where the server already has a Google or GitHub app for signing in, the matching integrations use it, and people click Connect instead of each organization creating an app. That approval is separate from signing in: signing in asks only for a name and an email, and connecting asks for the repositories, Drive or mail the integration needs. The steps for the person who hosts CrewHall are in Accounts.
Use my own OAuth app instead does the same with an app you create, for instance to show your company's name on the approval page.
With the per-person modes, a call uses the connection of whoever is asking (for a scheduled task, of whoever owns the task). An agent therefore never sees more than the person using it already could. Without a connection, the tool politely says so.
With a shared account, everyone who can use the agent effectively has that account's access. Choose what it can reach accordingly.
Which tools an agent gets
A server offers many tools; an agent gets only the ones an editor picks. In the agent's Configure tab, + next to Integrations adds the ones you tick, and asks what the agent may do there: Read only (the tools the server marks as only reading, the default), Everything, asking first (every tool, with the person approving each change), or Everything, without asking. Each integration then shows as one line with its buttons:
- Gear: which tools it may use. Everything that isn't read-only counts as a write and has to be ticked on purpose, with a warning: those are the calls that post the message, close the issue, change the document.
- Shield (once a write tool is ticked): whether the agent asks first. See below.
- Pause: the agent stops using it right away, without publishing, and keeps its settings for when you resume it.
- Bin: removes it, after you confirm.
The selection is part of the agent's published version, like everything else: adding, changing and removing take effect when you publish. Pausing and asking first are immediate.
Asking before writing
When an agent wants to use a write tool, it asks the person it works for first. The answer pauses, and a card shows what it wants to do: "Scout wants to post a message in Slack", with the channel and the text. Approve and the call runs; Reject, with a reason if you like, and the agent is told it wasn't done, and why. Nothing happens until you decide:
- in a chat, you have an hour; if you leave or reload the page, the card waits for you there;
- in a task's run, nobody is watching, so its owner gets a notification and has a day to answer in the task's conversation. Meanwhile the run waits, and the task page says so;
- an unanswered request expires and nothing is done; Stop cancels it too.
Only the person the agent acts for can decide. Every request and decision is in the audit log.
This is the default for every integration you add. To stop being asked, if you can edit the agent:
- click Always allow on the card: this change is approved, and the agent won't ask again before changing data in that integration;
- or click the shield on the integration's line: the agent then writes without asking, and the line says so in amber. Click it again to be asked first.
Agents that already used write tools before this existed keep working without asking until you turn it on.
Tokens and safety
- Tokens and sign-ins are encrypted at rest, tied to the row they belong to. They are never shown back to you, and never given to a model.
- A connection that stops working is marked as needing reconnection rather than failing silently.
- Results coming back from an integration are data, not instructions. A ticket description or a chat message can contain text aimed at your agent; agents are told as much, write tools are opt-in, and by default every write waits for your approval.
- Every call is in the audit log.
