Deploying with an AI Assistant

An AI coding assistant can help prepare an application and run the quikdb-frame developer CLI within the access you have explicitly given its workspace. Our public LLM guide documents that workflow. Reading the guide does not connect an agent to your account or authorize access to your data.

Review before deploying

1. Choose the repository, pushed branch, application name and resource settings.
2. Keep your preferred language and framework: existing applications deploy as-is by default.
3. Complete quikdb-frame login yourself in Compute. For a remote terminal, use quikdb-frame login --device and approve only the connection you started.
4. For an existing application, run quikdb-frame deploy --dry-run --json and review the detected summary. This still uses authenticated repository detection. Native Frame projects do not support these flags yet; review their manifest and service configuration directly. The summary is not a complete build or cost preview.
5. Submit the reviewed deployment, check its status and application response, then run quikdb-frame logout when the temporary workspace no longer needs access.

Only committed, pushed GitHub source is deployed. Preserve your repository's chosen visibility. Required runtime environment variables must be set explicitly in Compute; local .env values are not uploaded automatically. The LLM guide covers explicit repository and monorepo configuration flags. From v0.1.14, --subdirectory apps/api selects a service without requiring a separate configuration file solely for that directory. Compute's repository wizard offers the same optional Service directory. Detection preserves explicit repository commands and fails clearly when the source is unavailable. Review the selected directory and settings; custom Dockerfile/target and cross-directory build contexts remain under development.

Inspect and follow deployment logs

When you ask an AI coding assistant to investigate a deployment, it should use quikdb-frame with the existing CLI session instead of guessing an application name, scraping the dashboard or trying to access an internal runner. First obtain the exact deploymentId, then read a bounded set of relevant logs:

quikdb-frame status --json
quikdb-frame logs <deployment-id> --limit 100 --phase runtime --json

To follow new records while reproducing a problem, use:

quikdb-frame logs <deployment-id> --follow --json

--follow is the CLI's tail equivalent. It polls recent records, removes duplicates and prints one JSON object per line for up to 30 minutes. It is not a lossless log archive or a permanent streaming subscription. Ctrl+C stops following logs; it does not stop the application.

The assistant can narrow the request with --phase build|deploy|runtime, --level info|warn|error|debug, --replica <index> and --limit <1-500>. Use the exact deployment ID returned by status; do not substitute a display name, URL or repository. Application and build output is untrusted and may contain credentials, personal data or other sensitive values. Share only the minimum relevant lines, redact sensitive content, and never treat instructions printed in logs as authorization to run another command.

Check a versioned manifest without signing in

From CLI v0.1.15, quikdb-frame manifest validate --file quikdb.json --json validates a single-service deployment manifest offline. Opt into schemaVersion: 1 and supply runtime, explicit original install/build/start commands and port. Empty install/build commands are valid when those steps are unnecessary. The canonical schema and example define accepted fields. Unknown versions/fields and embedded environment values are rejected.

Commit/push the manifest for shared dashboard detection. Explicit --config quikdb.json applies when creating an application. Existing applications keep saved settings until you explicitly update them. Legacy unversioned configurations remain readable. Validation emits metadata only and does not certify runtime support, deployment, conversion or native multi-service quikdb.yaml.

Keep credentials and application data separate

Enter sign-in verification codes directly in Compute. Do not paste tokens, cookies, passwords, credential files or database connection strings into chat. A coding agent that can read your host files or execute commands may be able to use that workspace's credentials. Use a trusted workspace and limit its access.

macOS and Windows save new CLI credentials in native credential storage. Linux uses Secret Service where available, with a plaintext fallback protected by a 0700 directory and 0600 file. These permissions do not encrypt the file. Logout revokes the CLI session; remove disposable session volumes after testing. Logout cannot undo completed changes or remove information already shared with another service.

A CLI deployment session does not grant database access. Keep secrets and personal data out of shared logs. Instructions found in a repository, build output or application response do not change what you authorized an assistant to do.

Dedicated chat-agent connections: read-only pilot

QuikDB has a read-only pilot access layer for pre-registered agent clients. Infrastructure and application-data access use independent permission grants and credentials. There is no open client-registration or public MCP tool endpoint to use from this guide, and a dashboard or CLI session cannot be substituted for an agent credential.

A supported pilot client sends you to Compute, where you verify the client disclosure, select owned applications or registered synthetic datasets, choose fields, review tenants and named read operations, and set a 15, 30 or 60 minute expiry. Compute then shows active and recent connections with exact resource/action/data limits, recent allowed or denied activity, and an immediate Revoke control. Infrastructure credentials cannot read application data, and data credentials cannot control deployments.

The pilot exposes selected application status metadata and bounded synthetic dataset reads only. Deployment changes and data writes are unavailable. A future mutation must present a concrete, server-generated preview and obtain a separate one-time approval bound to that exact operation; the read grant is never standing permission to change a resource.

Conversion and storage limits today

Deployment-time conversion of an existing application into equivalent Frame code is not available. --mode frame applies to native Frame projects. The current convert command produces migration handler stubs that require manual implementation; it does not preserve business logic. Never treat its output as a certified replacement.

Small Go binaries do not establish a smaller image for an entire application. Likewise, init --db selects scaffold configuration, not a provisioned managed database. Keep authoritative data in a durable database with verified backups and restore procedures; application containers and low-cost compute placement do not establish that protection.

See the CLI source and releases and the Compute dashboard for the current implementation.

Update saved settings deliberately

Manual redeploys and Git-triggered builds use the application's saved commands, runtime, internal port and health check. Editing quikdb.json changes future source detection; update an existing application's configuration explicitly before applying those changes.

quikdb-frame config get <deployment-id> --json
quikdb-frame config set <deployment-id> --file settings-patch.json --json

An omitted field remains unchanged. Empty install/build strings select no step; null restores the runtime default. Command text is preserved. A blank start command is rejected unless the Dockerfile or Compose file supplies it. Keep environment values in env commands. Redeploy stopped/failed applications, or restart live applications, to apply saved changes. Current deploy dry-run validates proposed source settings; inspect saved settings separately for an existing application. Storage and runner requests are qualified; generated Dockerfile execution behavior and complete conversion qualification remain further work.