druks.toml for process and
deployment topology. Use the dashboard for operator choices that can change
without replacing the process.
The installer creates the deployment
.env from druks.toml. Compose, Druks,
and Drukbox consume this build artifact. Do not edit .env. Edit druks.toml,
then run the installer again to apply changes. druks setup creates .env but
does not restart services.
The file location determines its format. Repository files such as
.druks/software_factory/config.yml use YAML. Other repository dotfiles use
the same format. Operator files on the host use TOML because the installer must
create the environment without value changes. These files share no keys or
readers.
druks.toml is the authority for authored process configuration. Druks reserves
environment variables for infrastructure that Compose injects, such as database,
Redis, data, and container paths.
.env.example is the
host-run development template for that environment plane.
Deployment file
druks.toml has one table per operator concern:
A blank string means unset, and the renderer omits it from
.env. Use [env] for settings
without another druks.toml home, including additional DRUKS_* settings. The
renderer reports a key that it already owns as a configuration gap instead
of overriding its canonical value.
On a remote shape,
[sandbox.<provider>] accepts the variables documented by
Drukbox. Druks does not enumerate
providers. docker and exe select shape-specific first-write templates.
Every other provider name selects the generic remote shape. Drukbox validates it.
The local docker shape does not render [sandbox.<provider>]. Its Drukbox
service gets its environment from the defaults in deploy/compose.yaml.
The installer generates secrets only when it first creates the TOML. When you move or
recover an installation, preserve [secrets]. Use repeatable
druks setup ... --set key.path=value arguments for explicit scripted writes.
Core process settings
Postgres stores durable state. Redis does not store workflow state. It supports
short-lived concerns including webhook delivery claims, OAuth state and token
caches, and the sandbox provisioning gate.
Public URLs and access control
The
urls.webhook_host listener binds every interface. A second TLS
terminator on the same box, for example tailscale serve on the tailnet
address, collides with it on port 443. One of the two stops. To keep the
other addresses free, set DRUKS_WEBHOOK_BIND_HOST in [env] to the public
address. Caddy then serves only that address. To keep IPv6, list the IPv4
and the IPv6 addresses.
urls.endpoint and urls.webhook_host are different. The first is where an
operator’s browser reaches Druks. The second is the public ingress host for
webhook senders. They can share a hostname on exe.dev.
Druks does not authenticate browsers. Identity resolves per request, in this
order:
-
Personal access token. If an
Authorizationheader is present, it must authenticate. A malformed or inactive bearer returns a 401. Druks does not continue to another mode. -
Header mode (
header). The edge authenticates the operator. The edge can be exe.dev, Teleport, or Cloudflare Access. It supplies exactly one nonblankidentity.headervalue. Druks removes outer whitespace and maps the value to an account. Druks creates the account at first access. The edge controls who can access Druks. Account values are case-insensitive. -
JWT mode (
jwt). This mode uses the assertion channel fromheadermode, but its value is a signed JWT. Druks validates the RS256 signature againstidentity.jwks_url. It caches keys for five minutes and gets new keys after rotation. Theexp,iss, andaudclaims must match the configuration. Druks mapsidentity.jwt_identity_claimto an account. A validation error returns a 401 with the error class, not the token. Druks uses a fixed RS256 profile and does not negotiate it. -
No-authentication mode (
none). This mode has no authentication or identity edge. Druks resolves the only non-system account. Zero accounts is the setup state. The first completed harness connection creates the operator account from the provider-validated email. More than one non-system account is configuration drift. Druks refuses requests and startup in this state.
jwt mode, make sure that the edge uses the configured header,
claims, and rotation process.
The edge in header mode must authenticate each dashboard request. It must
remove client-supplied copies of the configured identity header. Then it must
put the authenticated value in the header. Otherwise, a client can select an
identity. The edge must also terminate TLS and set HSTS.
jwt mode has the same header-removal requirement. It also adds cryptographic
provenance. A forged value fails signature validation and returns a 401. Thus,
a bad proxy configuration does not create an impersonated identity.
The shipped Caddy listener is loopback HTTP behind that edge, and the Druks
web listener itself binds loopback by default. In none mode there is no
authentication. Keep the listener on loopback. Never publish it.
A public listener that bypasses the identity edge must not forward the
configured identity header. The shipped webhook listener serves only
provider-authenticated /_external/* and PAT-authenticated /mcp routes. These
routes do not resolve the header. A future public listener must keep the same
isolation.
Public POST /_external/* routes bypass the identity gate and use their own
authentication. Webhooks use signature validation. The notification response
route uses its correlation token. GET /api/auth/me answers without a
resolved account so the dashboard can render onboarding in the none-mode
setup state.
Personal access tokens
Agents and other non-browser clients use personal access tokens for the internal API. Mint these tokens in Settings → Tokens. Send a token asAuthorization: Bearer <token>. A token has the form
druks_pat_<prefix>_<secret>. Druks stores only the SHA-256 hash of the full
token. It shows the plaintext one time and expires the token after 365 days.
If the header is present, it must authenticate. A bad token returns a 401.
Druks does not use edge identity as a fallback. Token management accepts only a
signed-in identity. It refuses requests that contain an Authorization header.
Thus, a leaked token cannot mint or revoke tokens.
If someone compromises a token, mint a replacement. Then revoke the old token in
Settings → Tokens. Revocation is immediate. The list shows the prefix and last
use of each token.
Druks updates last use each hour. Agents consume the API
through the MCP endpoint. See
Connect your agent.
GitHub
Druks acts at GitHub as one operator GitHub App. This app is its service identity. The GitHub App receives webhooks and does domain writes such as branches, pull requests, comments, labels, and merges. Its credentials live encrypted in Postgres. They do not come from TOML, the environment, or a PEM file. Until an operator connects GitHub, agent runs stop with a direct message.druks doctor reports that no GitHub connection exists.
Connect it from Settings → Services. Create GitHub App starts the GitHub
manifest flow. Enter a GitHub organization, or leave the field empty for a
personal account. Accept the request on GitHub. Druks stores the credentials and
opens the installation page. Install the GitHub App on the applicable
repositories.
Before you create the app, set urls.endpoint to the dashboard base URL. If
you set urls.webhook_host, the webhook uses that host. Otherwise, it uses the
endpoint host.
You can paste the credentials of an existing GitHub App into the same card.
Enter the GitHub App ID, original PEM private key, and webhook secret. Druks
validates the credentials against GitHub and stores the app slug. Each operator
client then uses this service-identity row. Webhook deliveries use its stored
secret for validation.
To register the GitHub App manually, use this webhook URL:
Webhook URL:
https://<webhook-host>/_external/github/events/
Subscribe to issue comment, pull request, pull request review, and push events.
Install the GitHub App on the repositories that Druks will use. This
installation set defines where
software_factory can act. Personal access
tokens are not a supported substitute.
To upgrade an existing installation, paste the credentials one time on each
active host. Open Settings → Services. Connect GitHub with the existing operator
GitHub App ID, private key, and webhook secret. Do not create a replacement
GitHub App. The current webhook and installations continue to use the pasted
credentials.
Review identity (optional)
The bundledreview app can post its verdict reviews as a second
GitHub App, so GitHub accepts approvals on Druks-authored pull requests.
Configure it in Settings → Review. Enter the review GitHub App ID and its PEM
private key, both stored encrypted and empty-as-unset. Leave the pair empty and
reviews publish as operator comments. Set both values to publish separate
approval reviews. The review GitHub App needs read access to metadata and
contents, read/write access to pull requests, and no webhook.
GITHUB_API_URL defaults to https://api.github.com and can point every
client at another compatible GitHub API endpoint.
Ticketing integrations
Tracker credentials are service identities. Connect Linear or Jira Cloud from Settings → Services. The Linear identity uses an API key and webhook secret. The Jira identity uses a base URL, email, API token, and webhook secret. Druks validates the credentials before it stores them. Select the tracker and its workflow statuses in Settings → Software Factory. Webhook URLs remain/_external/linear/events/ and
/_external/jira/events/. The Jira webhook uses a Jira Automation
Send web request action.
Select Issue data (Jira format) as its body.
Druks accepts the REST issue JSON under issue. Put the shared token in the
x-druks-webhook-token header. druks doctor treats a disconnected tracker as
optional. It reports pending setup if the selected tracker lacks a connection.
Harnesses
Claude and Codex subscription credentials connect from Settings → Harnesses. The connection flow stores each credential in Postgres. Druks refreshes the credential on a schedule. It creates the CLI credential file inside each sandbox. It does not copy a host login. This is a capability connection for the requesting account — in a freshnone-mode install the first completed
connection also creates the operator account (see
access control).
Process settings such as DRUKS_CLAUDE_CONFIG_DIR and
DRUKS_CODEX_CONFIG_DIR point at optional non-auth CLI configuration to carry
into sandboxes. The Compose deployment mounts these read-only. Harness defaults
and per-agent model, effort, and timeout overrides live in dashboard settings.
A call refuses before provisioning a VM if its selected harness lacks a
connection.
Sandboxes
DRUKS_SANDBOX_KEYS_DIR remains a process environment override for the
per-host SSH private-key directory.
[sandbox].browser_login_proxy sends the browser login window through an
HTTP proxy. The login then leaves from a different IP than the box. Use it for
sign-in flows that refuse a login from the box IP. Only the login window uses the
proxy. Borrowed sessions keep the box IP. This is sufficient after Druks makes
the session.
If you do not set the proxy, the login uses the box IP. If you set the proxy and
the exit is not available, the login browser fails. It does not fall back to the
box IP.
The value can include a user name and password (http://user:pass@host:port).
The login browser authenticates the proxy. You do not need an external relay.
Druks does not run the exit. You supply the exit and set this value to it. There
are two common types.
Your own connection. Do these steps:
- Install Tailscale on a home device.
- Make the device an exit node in the Tailscale app.
- Add a
tailscale/tailscalecontainer to the deployment in userspace mode. - Set
TS_USERSPACE=true. - Set
TS_OUTBOUND_HTTP_PROXY_LISTEN=:8080. - Set
TS_EXTRA_ARGS=--exit-node=<your-device>. - Set
browser_login_proxy = http://172.17.0.1:8080.
browser_login_proxy = http://user:pass@isp-host:port.
An ISP IP passes the datacenter-ASN check. A detection service can still find it
and mark it as a proxy.
[sandbox].browser_login_tz sets the timezone of the login browser. Use an IANA
zone name, for example Europe/Madrid. The browser reports a region, and the IP
has a region.
Set both to the same region. Some sign-in flows compare these values. If the
two regions are different, a flow can refuse the login. Only the login window
uses this value. If you do not set it, the browser keeps the container default
timezone.
[sandbox].provider accepts any Drukbox provider name. docker selects the
local install shape, exe selects the exe.dev + tailnet shape, and every other
name selects the generic remote shape. Provider-specific credentials and host
options live in [sandbox.<provider>], and Drukbox interprets them. See
deployment or full local setup for the
topology.
Notifications
Manage destinations from the dashboard. The current destination type is a Slack incoming webhook. Actionable messages use Slack Block Kit. Other messages use the same URL through Apprise.SLACK_SIGNING_SECRET authenticates Slack interactivity callbacks.
Choose one enabled destination as the gate-notification destination in
Settings. A parked subjected run then produces a durable notification. Failure
to deliver the notification does not unpark or fail the run.
MCP servers
DRUKS_MCP_CATALOG points to a JSON catalog of server definitions. Druks loads
this catalog at startup. The packaged catalog contains an empty mcpServers
map. Thus, a new installation has no built-in servers. A deployment can point
DRUKS_MCP_CATALOG to a mounted file with its defaults.
Druks always loads a
catalog. A missing catalog stops startup. Catalogs contain definitions, not
tokens.
DRUKS_MCP_TRUSTED points to the trust-pins JSON for the official registry
badge. Druks calculates the badge. An entry is official if its reversed
publisher namespace matches the remote host. For example, com.grafana matches
*.grafana.com. A pin covers a value that this rule cannot derive. The value
shape selects one of two pin types:
- A publisher namespace (
"grafana": "io.github.grafana") identifies a publisher that the rule cannot match. The entry URL stays live from the registry. - An
http…URL ("sentry": "https://mcp.sentry.dev/mcp") supplies a hosted endpoint that the registry entry omits.
- A static token that Druks stores encrypted in Postgres
- A token from a named process environment variable
- An OAuth connection, which requires
urls.endpoint.
Skills
The dashboard installs skill collections from GitHub repositories.DRUKS_SKILLS_DIR selects the shared writable directory. Its default is
<DRUKS_DATA_DIR>/skills. A call receives the enabled skills that it requests.
If it requests none, it receives each enabled skill. A Software Factory build
requests the recommended set from the repository profile.
Druks excludes other
installed skills from the upload. The capability manifest records the delivered
set for each agent.
Credential custody and secrets at rest
secrets.secrets_key encrypts MCP tokens, OAuth grants, browser-session
payloads, and the GitHub service identity’s private key and webhook secret
with AES-256-GCM.
Each database column supplies authenticated associated data, and each value
gets a derived encryption key. The setting is one or more comma-separated,
base64-encoded 32-byte master keys:
druks.toml. Then run the installer again: