Connected agent apps
Connect a personal Slack or Plaud account, explicitly grant DreamLake's built-in agent access, and enable Apps in the composer for the turns that should use it. Existing project access and the general tool-write switch do not grant access to these accounts.
This feature requires the Connection v1 backend and configured provider OAuth clients. Until operator setup and a successful provider sign-in, the Apps screen shows setup required. Implementation tests use synthetic accounts and MCP fixtures; they are not evidence of a completed live provider connection.
Connect in the app
- Use
/go My connectionsto open Assistant connections, then find Apps for your agent. The canonical route is/<personal-namespace>/builtin-agent/connections. - Choose Slack or Plaud, review the access description and sign in through the provider.
- Select Allow agent reads, then Check connection. A successful check verifies sign-in and supported tool discovery; it reads no account content.
- Enable Apps in the composer and ask a question, such as “Summarize my recent Plaud meeting and find related Slack discussion.”
Slack read consent includes accessible public/private conversations, direct messages and files. This version has no channel allowlist: use an account whose provider permissions match the intended scope. Plaud reads synced recordings, transcripts and existing summaries; recordings need Plaud cloud sync. No Plaud editing or deletion is exposed.
For Slack messages, separately select Authorize Slack sending and Enable reviewed sends. The agent can then prepare a proposal. Review the account, exact destination, thread and message in Apps, and select Send this exact message. Preparing a proposal does not send it. Sending consent is separate from read access.
CLI and tools
The CLI reuses server-owned Connection OAuth. It never receives provider access or refresh tokens.
Use a personal --namespace if the active namespace is an organization. Discover current tool argument schemas before constructing the JSON object. Inputs are limited to 16 KiB.
CLI confirmation displays the exact action and defaults to declining. It requires an interactive terminal and rejects --json; there is no unattended --yes. The TypeScript client exposes the same app APIs for callers that obtain human authorization themselves. See the CLI app reference for the command contract.
The model receives four tools:
| Tool | Behavior |
|---|---|
dreamlake_apps_list | List personal app connections and explicit grants. |
dreamlake_apps_tools | Discover reviewed tool names and argument schemas. |
dreamlake_apps_read | Execute an allowed read tool. |
dreamlake_apps_prepare_action | Save an encrypted Slack proposal and return its review link. |
There is no model tool to grant access, sign in, or confirm sending. Generic resource tools exclude the /agent API family, so they cannot invoke app confirmation indirectly. App content and provider descriptions are untrusted data, not instructions or authorization.
Receipts, revocation and recovery
A proposal expires after ten minutes. Confirmation is bound to its owner, namespace, connection, grant revision, connection revision, credential version, tool schema and payload hash. A database claim allows at most one dispatch. Changed access or schema invalidates the proposal.
A timeout, provider error after dispatch, or lost completion record produces an uncertain outcome. Check the saved receipt and Slack before preparing another proposal. There is no automatic write retry. An interrupted process can leave a sending record; after a minute the review interface reports it as uncertain. Terminal receipts remain durable and are stored encrypted; no automatic receipt retention cleanup is configured in this version.
Revoking a grant stops future access and suppresses a read result if access changes while the request runs. Disable/delete the underlying account through Connections to stop all its consumers. Content already saved in a conversation or memo remains there. The catalog displays up to 100 accounts and the 20 most recent actions; known action IDs remain reviewable directly.
Operator setup
Use the existing Connection v1 OAuth, encryption keyring, secret store, callback configuration and renewal infrastructure. Configure clients through deployment environment secrets, not a UI text field or command-line token argument. Both deployment workflows pass the four *_MCP_CLIENT_* secrets to the server. Missing configuration keeps that provider in setup required. These clients are separate from the native Slack login client; do not reuse its tokens or assume its registered callback covers MCP.
| Provider | Fixed MCP endpoint | Required deployment configuration |
|---|---|---|
| Slack | https://mcp.slack.com/mcp | SLACK_MCP_CLIENT_ID, SLACK_MCP_CLIENT_SECRET |
| Plaud | https://mcp.plaud.ai/mcp | PLAUD_MCP_CLIENT_ID; PLAUD_MCP_CLIENT_SECRET if issued for the registered client |
Register each exact callback under the deployment's configured public API origin:
/api/connections/oauth/slack-mcp/callback/api/connections/oauth/plaud-mcp/callback
Slack requires an eligible registered application; it does not support dynamic client registration for this MCP flow. Plaud advertises registration metadata, but this implementation expects an operator-registered client and does not auto-register one for each user. Clients must support S256 PKCE and the configured callback. Plaud clients must accept client_secret_post or public-client token requests; client_secret_basic-only clients are not supported here.
Consult Slack's official MCP documentation and Plaud's official MCP documentation for provider eligibility and registration. Provider configuration and current accounts must be verified before claiming live readiness.
Slack read authorization requests search:read.public, search:read.private, search:read.im, search:read.mpim, search:read.files, search:read.users, channels:history, groups:history, im:history, mpim:history, channels:read, groups:read, im:read, mpim:read, files:read, users:read, and canvases:read. The read-write profile adds chat:write. Actual returned consent is checked again before writes. If a refresh response omits scope, sending fails closed and requires renewed consent.
Deploy the backend before UI/CLI consumers. Generate the Prisma client during the build. In production, use the reviewed-main Schema indexes (manual, scoped) workflow with scope builtin-agent: inspect its read-only plan, then apply its exact database-bound plan hash. This scope includes the unique AgentAppGrant(userId, connectionId) index, the AgentAppGrant(namespaceId, userId) lookup index, the AgentAppAction(userId, createdAt) receipt index, and existing built-in-agent indexes. Review every proposed addition before applying. Production deployment refuses to create missing indexes as a side effect; the Connection startup gate also verifies them before serving traffic. pnpm db:ensure-indexes is for a separately verified local/development database, not a substitute for the production plan/apply workflow.
Supported transport is Streamable HTTP to the fixed provider endpoint, with redirects rejected, a 25-second session budget, bounded pagination, and a 256-KiB response cap. Tool names are statically reviewed. Provider schemas with remote references, regular-expression validation, excessive depth or size are excluded. Unknown tools do not become available merely because discovery returns them.
The live acceptance sequence is: provision clients and callbacks; sign in to a test account; verify discovery; run one read; revoke and confirm refusal; reconnect with explicit Slack sending consent; prepare and inspect one message in an authorized test channel; confirm once and verify its receipt and provider result. These final provider checks require the registered clients and consenting accounts.