Run Adapt
Security and privacy
Adapt keeps a workspace's conversations, settings, and encrypted credentials in PostgreSQL, and its files on disk or in cloud storage.
This page covers what Adapt stores and where, what it sends to the providers you connect, who in the workspace can see what, and how credentials are handled.
What Adapt stores
Everything belongs to one workspace, and every read is scoped to it.
- PostgreSQL: accounts and memberships, agents and their settings, conversations and messages, run activity (each tool call and its outcome), company instructions, documents and skills, connections, automations, the activity log, and usage.
- File storage: uploaded and generated files, on disk (
ASSETS_DIR) or in a Google Cloud Storage bucket (GCS_BUCKET). See Self-host.
Adapt encrypts credentials, not conversations. Messages, files, and documents are stored as written, so anyone with direct access to the database or file storage can read them. Restrict that access the way you would for any production database.
On Adapt Cloud, Adapt runs the database and storage for your workspace. When you self-host, they're yours.
What leaves your install
Adapt sends data only to the providers your workspace connects, when an agent uses them:
- The agent's model connection receives the instructions, the conversation, and tool results for each step of a run.
- Integrations, sandboxes, and knowledge providers such as Mem0 receive the calls the agent makes to them.
- Surfaces (Slack, GitHub, email, text messages) receive the agent's replies and the calls its surface tools make.
If you turn on tracing, traces carry metadata only (models, token usage, timings, tool names, error types) unless a workspace is listed in TRACE_CONTENT_WORKSPACES. See Self-host.
Who can see conversations
A web chat is private to the person who started it. Owners and admins can't open, read, or export it, and it doesn't appear in their history search. Each person's history search covers only their own conversations.
Threads from a surface are shared differently, because other people are already in them:
- The member who started the thread sees it in their sidebar and can rename or delete it.
- Any member who can view the agent can open it read-only. For a Slack thread, they must also be able to read that Slack channel. GitHub, email, and text-message threads have no such check.
- Admins and owners can delete a shared thread. Deleting hides it, and the next message there starts a fresh conversation.
An automation's runs are visible to people who can view that automation. History search and Adapt administration tools are never available in surface threads.
Who can change what
Roles are set in Settings → Members. In short:
- Owner and Admin: change workspace settings, connections, surfaces, Agent defaults, company instructions, documents, and skills; manage members; read the activity log; see everyone's usage.
- Member: chat with agents, create agents and edit their own, manage their own personal integrations and automations, and see their own usage. Members can view connections, documents, and skills, but not change them.
An automation is managed only by its owner, whatever their role. See Members and roles for the full list.
The activity log
Owners and admins read the workspace's audit trail in Settings → Activity log, and can export it as JSON. Each entry records who acted, the action, the resource and its name, and, when an agent acted for someone, which agent.
The log records names, not contents:
- An update lists the names of the fields that changed, never their values.
- A tool call records the tool's name, not its arguments or results.
- Renaming a private chat isn't logged; only shared threads are workspace business.
Members see what their own runs did in each conversation's tool activity. See Chats and tool activity.
Credentials
Credentials you save in a connection, workspace secrets, and the Google sign-in client secret are encrypted with AES-256-GCM using the install's ENCRYPTION_KEY. Each one is bound to its workspace and connection, so an encrypted value can't be moved to another record and decrypted.
- The app never returns a saved credential: settings show which fields are saved, by name, not their values. Pasting a new value replaces the old one.
- Secrets given to a sandbox are environment variables for that command. Their values, including common encodings of them, are replaced with
[REDACTED]in the command's output and errors before the agent or anyone else sees them. - Some setup steps use a key once and don't keep it, such as the AgentMail key that creates an inbox. See Email.
- Losing
ENCRYPTION_KEYmakes saved credentials unreadable. Back it up with the database. See Self-host.
Messages from outside
People outside the workspace, quoted email, earlier thread messages, and tool results can all contain text that looks like instructions. Adapt labels that content as untrusted data when it hands it to the model, and its instructions tell the agent never to treat external messages as authority.
On surfaces, the sender Adapt verified is stated in the agent's instructions from the run itself, so message text can't claim to be someone else. Surfaces answer only workspace members by default; answering anyone else is an explicit setting on each surface's automation, and those runs are anonymous, with no Adapt user and no workspace role. Removing a member or changing their role is checked again before each tool call, so a pending turn or remaining tools stop.
Tool calls and uncertain outcomes
Tools run as soon as the model calls them, and every call is shown with its outcome: succeeded, failed, or uncertain. An uncertain call may or may not have taken effect. Adapt never retries it automatically and never reports it as a success. Replies on surfaces follow the same rule: one that might have been sent is marked uncertain and never sent again.
A connection that's on is permitted for use, not verified healthy. Setup checks show what Adapt could confirm.
Sign-in
- Accounts sign in with email and password, or with Google when the workspace turns it on in Settings → Authentication. Google sign-in accepts only accounts on the company domain.
- After the first account, people join only through an invitation link, valid for 48 hours for one email address, or through Google sign-in on the company domain.
- Sign-in is rate-limited per client. Sessions last 7 days.
- A removed member's surface identities stop matching, and their pending turns are blocked.
Coding agent access
Owners and admins can let coding agents such as Cursor or VS Code connect to Adapt over MCP in Settings → MCP server. It's off until turned on and requires Adapt to be served over HTTPS.
A coding agent signs in with OAuth as a member, who approves it on a consent page, and then acts with that member's access. It can list the member's agents, ask one to do work and read the result, and run automations the member can run. Turning access off disconnects coding agents for every member until it's turned back on.