Set up
Secrets
A secret is a named, write-only value, such as an API key, that Adapt passes into an agent's sandbox as an environment variable. Admins add secrets once; each agent gets only the ones selected for it.
Secrets are for code an agent runs in its sandbox. Credentials for models, sandboxes, knowledge and integrations belong to their connections instead.
Add secrets
Owners and admins manage secrets in Settings → Secrets. Add a row for each secret with a name, a value and an optional description, or paste the contents of a .env file:
STRIPE_API_KEY=sk_live_...
WAREHOUSE_URL=postgres://reporting@warehouse.internal/analytics
Pasted lines merge with the list by name, and the last value for a name wins. Adapt reads the text as plain assignments; it doesn't run it through a shell. Choose Save changes to store the whole list at once.
After saving, a value is never shown again. Members see each secret's name and description so they can pick it for their agents; no one can read a stored value back through Adapt.
- Names: letters, digits and underscores, not starting with a digit, up to 128 characters.
- Values: up to 32 KiB each, and up to 1 MiB for all values together.
- Descriptions: up to 2,000 characters.
- Up to 256 secrets per workspace.
Reserved names
Some names would override how the sandbox runs or how Adapt signs in to GitHub, so they can't be used:
- Any name starting with
ADAPT_,_ADAPT_,LD_,DYLD_,NODE_,BASH_FUNC_,GIT_,SSH_,DOCKER_,PYTHONorNPM_CONFIG_. - Shell and system variables such as
PATH,HOME,SHELL,USER,TMPDIR,PWDandIFS. - GitHub credentials and hosts:
GITHUB_TOKEN,GH_TOKEN,GITHUB_ENTERPRISE_TOKEN,GH_ENTERPRISE_TOKEN,GH_HOST,GITHUB_API_URL,GITHUB_SERVER_URLand relatedGH_settings. - Proxy and certificate settings:
HTTP_PROXY,HTTPS_PROXY,ALL_PROXY,NO_PROXY,SSL_CERT_FILE,SSL_CERT_DIR,CURL_CA_BUNDLE,REQUESTS_CA_BUNDLE. - Language runtime options such as
PERL5OPT,RUBYOPTandJAVA_TOOL_OPTIONS.
Names are checked without regard to case.
Give an agent secrets
Open the agent's sandbox settings and choose its secrets under Secrets. Only the secrets selected there reach that agent's sandbox. Members can select secrets for agents they own; owners and admins can select them for any agent.
A few rules keep the selection deliberate:
- Agent defaults can't select secrets. Every new agent starts with none.
- An agent can't add secrets to itself or another agent from a chat, even when it can manage Adapt. Choose them in the agent's settings.
- Exported agent files leave out secret selections.
Adapt reads the selection again before each sandbox operation, so a change applies on the next command. A conversation keeps its sandbox only while the selected secrets and their values stay the same. If you rotate a value, remove a secret or change an agent's selection, the conversation's next command runs in a new sandbox. Editing only a description doesn't replace the sandbox.
Update and delete secrets
Leave a saved value unchanged to keep it, or type a new one to replace it. Removing a row and saving deletes the secret and removes it from every agent that selected it.
Saves are checked against the list you loaded. If someone else saved in the meantime, the save is refused with nothing changed; reload and make your edits again.
How values are protected
- Values are encrypted at rest with AES-256-GCM, bound to the workspace and the secret, using the server's
ENCRYPTION_KEY. - They're decrypted only to create or run commands in a sandbox for an agent that selected them.
- Known values are replaced with
[REDACTED]in command output, file reads and provider errors that come back to the agent and the chat, including their JSON-escaped, base64 and URL-encoded forms. - The activity log records that secrets changed, never their values.
Redaction stops accidental leaks, not deliberate ones: code running in the sandbox can read the values and send them elsewhere. Select a secret only for agents whose tools, sandbox provider and container image you trust with it.