Agents

Every agent works on a workspace the same way: through Dataembed's MCP server, with the tools its scope and its person's role allow, logged under its name, undoable where the change is. That includes the copilot.

The copilotCopy Anchor Link

⌘J / Ctrl+J opens it in the side panel. It sees what's on the page (the object, the view, the selected records, the open record), and answers from tool results only.

• It runs on Claude Managed Agents, as Claude Opus 5.5, with web search, web fetch and a sandbox for code. It reaches Dataembed only over MCP, as you: your role, now.
• Undo first. It makes the change you asked for at once rather than asking first; every change to records, views, layouts, plugins, objects and fields is in the activity, ready to undo.
• What can't be undone waits for an approval (below). Some web or sandbox steps ask you first: Allow or deny.
• It lists your threads, where you start a new one or delete one. Threads are yours alone: nobody else, owner or agent, reads them.
• A thread has a budget ($5). Past it, the thread stops; start a new one.
• It needs AI set up on the deployment: until then it says so, and the plugin shows Needs configuration.

Copilot instructions give it, and every connected agent, standing notes about the workspace.

JobsCopy Anchor Link

Long work runs unattended as a job with its own budget: Fill with AI fills a field for many records, Find records adds records from the web. A job acts as the person who started it, and shows on the object's toolbar with its progress and Cancel (its starter, or an owner or admin). Agents follow it with jobs_get and stop it with jobs_cancel.

ApprovalsCopy Anchor Link

A change that can't be undone (deleting records, fields, plugins or a data source, a schema change, an object's key, a connection, API keys, the team, the workspace's name) doesn't run when an agent calls it, whichever agent and whatever its scope. It's held:

• the call fails with PENDING_APPROVAL (HTTP 403) and an approval id, summary and expiry;
• any owner or admin approves or rejects it: in Settings → AI & agents → Approvals, or with Approve and run / Reject on the copilot's card;
• once approved, it runs as the agent that asked; only that agent reads what it answered (approvals_get);
• unanswered, it expires after 24 hours.

A person signed in does these themselves, after a confirmation, without an approval.

Bring your own agentCopy Anchor Link

Claude, ChatGPT, Claude Code, Cursor and other MCP clients connect to https://<your-dataembed>/api/mcp. Settings → AI & agents shows the URL and ready-to-copy snippets.

Sign in (OAuth)Copy Anchor Link

Add the URL as a custom connector or MCP server. Claude Code:

claude mcp add --transport http dataembed https://<your-dataembed>/api/mcp

The app sends you to Dataembed to sign in, pick the workspace and approve what it may do: one scope, starting at the highest it asked for below admin (only owners and admins are offered admin). An app that names a verified domain shows it; one that registered itself shows Unverified app. It then acts as you, with your role now; its changes show as "Claude via Ada". Access tokens last an hour and refresh for 30 days; an app that uses DPoP gets tokens bound to its key.

Settings → AI & agents → Connected apps lists the apps (owners and admins see everyone's); Disconnect stops one at once. Leaving the workspace disconnects yours there.

Or an API keyCopy Anchor Link

For clients that only send a header, and scripts. Owners and admins Create API key in Settings → AI & agents: a name, a scope (write by default), an optional expiry. The key (de_…) is shown once. It acts for its creator with their role now; one key per agent, so each can be revoked alone.

claude mcp add --transport http dataembed https://<your-dataembed>/api/mcp \
  --header "Authorization: Bearer de_your_api_key"

{
  "mcpServers": {
    "dataembed": {
      "headers": { "Authorization": "Bearer de_your_api_key" },
      "url": "https://<your-dataembed>/api/mcp"
    }
  }
}

The same key calls the JSON API. A good first prompt: "Describe my dataembed workspace." On connecting, the server tells the agent what it can't do and where a person does it, which tools its scope or its person's role holds back, and the workspace's instructions.

Settings → AI & agents lists each key with its state (status in api_keys_list): active, paused while its creator is suspended, dead once they leave or it has no scope, expired or revoked. A key made by a key, an app or an agent (the copilot, AI fill, web search) names who issued it. Revoking stops an active or paused key at once, with every key issued through it; any other is left as it is.

ScopesCopy Anchor Link

A key or an app has one scope, and can do only what it allows and its person's role allows now. Each includes the one before. admin is opt-in, and only owners and admins grant it.

ToolsCopy Anchor Link

One tool per operation of the API, named after it (records_update_many is records.updateMany), listed only when the credential's scope reaches it and its person's role allows it: 75 for an owner with admin, 74 for an admin, 45 with write, 24 with read.

• Tools name objects and fields by slug, and return each record with its label and a link. Start with workspace_describe.
• records_list pages with filters (total.kind says whether its count is exact, an estimate or at_least); records_aggregate counts and sums per group; sql_query runs one read-only statement (Query).
• records_import takes up to 500 rows a call; run it with dryRun: true first.
• schema_preview shows the SQL schema_change would run.
• views_update and layouts_update change things for the person only; views_save_for_everyone and layouts_save_for_everyone for everyone.
• plugins_manifest_schema, then plugins_create / plugins_update, build workspace plugins; plugins_run_action runs a bundled plugin's action, which may start a job.
• Every write returns the activityIds that activity_undo takes.

What agents can't doCopy Anchor Link

These are for people signed in, in the app: they act outside one workspace, widen access, or are a person's own. They're no tool, and the API refuses them to keys.

No escalationCopy Anchor Link

• An agent never grants more than it holds: a key made by a key gets no higher scope, expires no later and is revoked with it.
• Nobody changes their own role, and only an owner makes or changes owners.
• The current role always applies: demote a key's creator and its admin tools stop at once.
• /api/mcp and /api/v1 refuse browser requests from other sites; each credential makes at most 300 requests a minute.