Direct the work in Studio
Use Drever Studio to answer adaptive questions, approve the content Storyboard, follow agent work, and direct one slide or the whole live draft.
From the first question to a live deck.
Shape the intent, approve the story, then direct the real presentation in the same room.
Keep the creation loop in one room
Studio is Drever's local, development-only workspace for directing a coding agent. It keeps the brief, adaptive questions, content Storyboard, live draft, agent activity, approvals, and feedback beside the presentation without turning the browser into a hosted editor.
Start it from the project root:
npm run dev -- --open studio --agent codexReplace codex with another supported adapter when needed. Drever opens the exact loopback-only
/studio URL on an ordinary desktop. Headless environments print the URL without attempting to
open a browser.
The Creation room belongs to that running development command. Keep its terminal or supervised
agent task alive while you review and revise the deck; ending the parent process also ends Studio.
The terminal prints deduplicated Studio action and phase changes even when Vite itself is quiet.
It preserves the exact Creation room URL in the scrollback. Closing the last Studio tab prints the
address again immediately, and later semantic phase changes keep it near the latest output while no
Creation room is open, so you can reopen it without restarting the development command.
Coding-agent supervisors use those lines—or drever studio status --json—as the source of truth;
they do not require a chat continue message to keep an already active managed agent working.
When the original request already names the subject, pass it once as --topic "React 19 changes".
Studio prefills the Brief but still waits for the user to confirm it; if no topic was supplied, open
Studio directly and collect it there instead of asking once in chat and again in the browser.
Studio does not call a model by itself. The selected CLI owns its account, credentials, model, and provider policy. Drever normalizes only bounded public work summaries, plan state, semantic tool events, and supported actions before sending them to the browser.
Move through four deliberate phases
Studio follows the same project contract as the agent's installed Drever workflows:
- Brief — state the topic, audience, desired outcome, duration, visible density, and motion intensity. Free-form duration and direction remain available when the suggested choices do not fit.
- Direction — answer a small set of questions chosen from the subject and earlier answers. Every round can accept free-form detail or Skip remaining questions — surprise me.
- Storyboard — review the content and sequence before per-slide layout or motion exists. The
agent publishes a complete
drever.plan.json, then waits for explicit approval or edits. - Draft — inspect the real Audience View, speaker Notes, and public agent activity while the agent refines one coherent draft.
Completed stages remain clickable in Studio's progress header. You can return to Brief or Direction to review the saved choices without restarting work. If you change one, Studio confirms the rebuild, marks its dependent Storyboard and Draft as superseded, and keeps those previous artifacts available read-only until their replacements are ready.
The Storyboard is intentionally content-first. It shows each slide's job, purpose, evidence, and anchor artifact, but defers per-slide layout and motion until the story is approved. This keeps the first checkpoint fast and prevents styling from making a weak argument feel finished.
If Studio is unavailable, the installed workflow asks the same questions in chat and opens the
read-only /storyboard route for approval. The persisted files and approval boundary are identical.
Direct the live draft without restarting it
After approval, the agent publishes a complete, valid Draft 1 atomically. It writes the whole content draft, runs the source check once, and exposes that stable preview before optional research, custom art direction, or motion refinement. Studio never treats a token-streamed fragment of MDX as a presentation, but it also does not hold the first useful preview until final polish is complete.
Use the slide rail to select the current preview. Feedback has an independent scope:
- This slide sends a focused request for the selected slide.
- Entire deck sends one request about the story or visual system without changing the previewed slide.
When the live draft is ready, Find improvements asks the connected agent for a short, analysis-only review. Studio shows at most three evidence-based ideas across story, visual, motion, and accessibility concerns. Selecting one only places its proposal into the feedback box and selects its scope; nothing changes until you explicitly send that feedback to the agent.
The agent updates the same ordinary project files and the same live preview. slides.mdx,
brief.md, drever.plan.json, local design files, assets, configuration, and Git remain the source
of truth. A revision does not create a hidden Studio document.
After Send to agent acknowledges a request, Studio keeps a prominent accepted, reconnecting, or refining state visible until that exact action settles. The latest valid draft remains available to inspect during the change. Studio does not expose private chain-of-thought, raw commands, tool payloads, or provider transcripts; it shows one bounded public work summary instead.
Managed Codex keeps workspace-write sandboxing and on-request approvals. Eligible low-risk local
checks go through Codex Auto-review,
while provider requests that still surface a user decision keep Studio's prominent approval action.
Connect the agent you already use
Codex and Claude Code use dedicated native transports. Agents with a compatible Agent Client Protocol server share one ACP adapter. Other coding agents can use the portable action-journal bridge when they can run project commands but expose no stable interactive protocol.
| Agent | Studio transport | Approval path |
|---|---|---|
| Codex | Native App Server | Studio |
| Claude Code | Native streaming CLI | Claude permission policy |
| Gemini CLI | ACP v1 | Studio when advertised by the agent |
| GitHub Copilot CLI | ACP v1 public preview | Studio when advertised by the agent |
| Goose | ACP v1 | Studio when advertised by the agent |
| Cursor CLI | ACP v1 | Studio when advertised by the agent |
| OpenCode | ACP v1 | Studio when advertised by the agent |
| OpenHands | ACP v1 | Studio when advertised by the agent |
| Cline | ACP v1 | Studio when advertised by the agent |
Start a managed adapter only when the matching CLI is installed and authenticated:
npm run dev -- --open studio --agent claudeWithout --agent, Studio uses the provider-neutral bridge to the coding-agent task that started
the workflow. Opening /studio alone does not launch or wake a model. If the task stops polling,
submitted actions remain queued until it resumes, and Studio reports that the connection is absent
instead of pretending work is active.
If a managed adapter disconnects while an action is pending, Studio first restarts the local provider and replays that same journal revision once; it does not append duplicate brief or feedback instructions. A reconnect control remains available if that bounded automatic recovery also fails.
Know what stays local
Studio is served only by drever dev. It is excluded from production builds, PDF export, and the
audience bundle. Session files under .drever/studio are ephemeral coordination state and should
not be committed.
The browser receives no API key and no arbitrary file access. Provider credentials and protocol details remain on the local server. Drever adds no telemetry to the generated project; the chosen AI provider, remote media, and eventual hosting platform keep their own privacy contracts.
When the draft is ready, continue with Rendered review for deterministic layout evidence before delivery.
Understand the project-local workflows, context, and read-only connections that keep agent work grounded.