Work in the room with clarity, control, and a visible trail.

Getting started

1. Sign in. Use the email sign-in link first so the same account works in browser and desktop. If you still need recovery-code sign-in, open the recovery login page from the access screen.
2. Create or open a project. Give the room a short name that identifies the outcome.
3. Set the objective. Describe the result, constraints, and what success looks like.
4. Add context. Attach the relevant small files and open the real project URL in the Browser pane.
5. Send a focused request. Select a mode or agent, then review the response, tasks, files, and reviews.

Understand the workspace

Chat

The central timeline shows your messages, agent decisions, next steps, and project preview cards. It is the main record of the project conversation.

Tasks

Tasks turn agreed work into trackable items. Update their status as work moves from pending to complete.

Reviews

Material decisions stay visible until you approve or reject them. Opening a review only opens its details; it never records a decision.

Agents and insights

Agents shows provider runs and status. Insights combines summary statistics with a complete readable project report, decisions, findings, progress, and deliverables.

Choose how the team works

Balanced is the recommended default. It considers planning, building, and review. Use the agent chips when you want a specific perspective:

• GPT · Strategist: scope, planning, and synthesis.
• Baki · Researcher: research and supporting context.
• Codex · Builder: implementation and technical execution.
• Claude · Reviewer: critique, risk, and quality review.

Auto-route lets the orchestrator select the best specialist. Turn it off when you want manual control. The actual models available depend on the server-side providers configured by the administrator.

Use the Browser pane

1. Enter an https:// or local development URL in the address field.
2. Use refresh after the project changes.
3. Switch desktop, tablet, and mobile sizes to check responsiveness.
4. Expand the pane for close inspection or open the URL externally.

Production previews are sandboxed. Do not enter passwords or sensitive customer data into an unfamiliar page loaded in the preview.

Files and tasks

Use Attach to add focused context. Supported file types are plain text, Markdown, CSV, JSON, PDF, PNG, JPEG, and WebP. Each file is limited to 2 MB. Large repositories and secrets should stay in approved development systems, not browser uploads.

Tasks can be created from the workspace and associated with an agent. Keep titles action-oriented and small enough to verify.

Build and run automation

1. Open Automation. Choose the Automation tab or the Automation item in the sidebar.
2. Create a saved automation. Give it a clear name, description, trigger, and retry limit.
3. Add ordered steps. Agent work produces a Markdown deliverable, Create task adds a project task, and Review checkpoint pauses until an owner decides.
4. Choose a trigger. Manual runs start with Run now. Schedules run at the saved interval, with a minimum of five minutes. Secure webhooks display a secret URL once; copy it into an approved secret manager.
5. Monitor the run. Open Latest run to view step state, attempts, outputs, errors, and the execution log.
6. Stay in control. Pause stops before the next step, Resume continues, Cancel ends unfinished work, and Retry failed step restarts only the failed point.

Limits and notifications

Each project can save up to 25 active automations. An account can have up to three active runs at once, each automation can contain 1–20 steps, and retries are limited to five. Agent steps consume Project Credits. The notification area reports reviews, failures, completed runs, skipped schedules, and low credits.

Background execution

Runs continue through the server-side job queue if you leave the Automation tab. In production, PostgreSQL stores run state and the worker leases each job so multiple server instances cannot execute the same queued step simultaneously.

Review before deciding

1. Open the Reviews tab. Each row shows its title, short description, requester, and current status.
2. Select Open review to read the full decision request. Opening it does not approve anything.
3. Compare the request with the project objective and safety checklist.
4. Read the related agent activity and download any attached deliverables.
5. Choose Approve decision only when the evidence is sufficient. Choose Reject when more work or clarification is required.

Completed decisions remain listed with their recorded status for accountability. Automation review checkpoints also appear in the execution view and continue only after an explicit review.

Read and download insights

The statistics at the top of Insights are only the summary. The full report underneath contains the current objective, task progress, review decisions, agent findings and activity, and downloadable files and deliverables.

1. Open Insights from the project tabs.
2. Review the completion, review, and agent-run statistics.
3. Read the complete report sections underneath; they are generated from the current saved project record.
4. Use Download report for a readable Markdown copy.
5. Use Download project archive for the project ZIP containing the report, manifest, uploaded files, and agent outputs.

Settings sections

The Settings screen now saves operational project preferences to PostgreSQL, applies ownership checks, and records audit events. Live sections include account profile, AI model selection, agent roles, automation defaults, notifications, project defaults, browser preview, security and privacy, data export and retention, integrations and webhooks, team access, and help links.

• Account & profile
• AI providers & models
• Agent roles
• Automation defaults
• Notifications
• Project defaults
• Browser & preview
• Security & privacy
• Usage & billing (read-only until live billing is approved)
• Data, export & retention
• Team & access
• Integrations & webhooks
• Desktop app (unavailable until the signed package ships)
• Help & support

Provider secrets stay server-side. Sensitive updates require confirmation, and unavailable sections are clearly labeled instead of pretending to work.

Security and privacy

• Never share your recovery code, sign-in link, session token, provider key, or administrator token.
• Provider keys belong only in server secrets—never in the browser or a VITE_* variable.
• Sign out on shared devices.
• Use only project data you are authorized to process.
• Report unexpected account access through the support form below.

Each project is isolated to its claimed account. Browser sessions use HTTP-only cookies; desktop/API clients may use bearer tokens over HTTPS.

Troubleshooting

The browser pane is blank

Confirm the URL is reachable and uses HTTPS in production. Refresh once. If the website blocks embedding, open it externally.

An AI run does not start

Check the Agents tab for its status. The administrator may need to configure the AI provider or the account may have reached its plan limit.

An automation run is retrying or failed

Open Latest run and inspect the failed step and execution log. Check provider configuration and Project Credits. Use Retry failed step only after correcting the cause; completed steps are preserved.

A scheduled automation did not run

Confirm the automation is Active, its next-run time is in the future, and the account has an active session and available capacity. A skipped schedule creates a notification with the reason.

A file will not upload

Confirm it is a supported type and no larger than 2 MB. Remove secrets before uploading.

I cannot sign in

Request a fresh sign-in link first. If you still need recovery-code sign-in, open the recovery login page from the access screen and use the newest access email.

Changes are not visible

Refresh the workspace and then the Browser pane. Confirm the browser points to the current project server, not an older v1.1 process or URL.

Frequently asked questions

Contact support

Use the authenticated support form below. Sign in first so the request can be securely associated with your account. Include what you expected, what happened, and the exact step where it stopped.