Turn ideas into results
The chat workspace is where Lévin™ Harness comes together. Describe what you want to accomplish, and Levin brings project files, molecular structures, tools, and compute into one working session.
The chat workspace has three parts: the sidebar organizes projects and conversations, the transcript records the work, and the composer controls what Levin receives and what it may do. You do not need to configure every control before you begin. First choose the right conversation scope, then add only the context the task needs.
Input and controls
The composer is where a task begins. Describe what to build, fix, or analyze, then add only the context it needs through the + menu, @, #, or attachments. Project files, Goal mode, the Agent model, thinking strength, and tool permissions are collected here so you can review what the message includes and what Levin may do before you send it. Once a conversation starts, its model stays stable; you can still adjust permissions as the scope or risk changes, and queue or steer follow-up messages while it runs.
Control toolbox
1. Context management
Opens the + menu to add capabilities, structures, attachments, a temporary chat, or a planning pass.
Use it when: the message needs context beyond the written request.
2. Set a goal
Turns the request into a managed Goal that Levin can advance across stages while tracking progress.
Use it when: Levin should manage a long-running task and continue after waiting or interruption.
3. Project files
Browses the active project and brings the files Levin should work with into the conversation.
Use it when: Levin needs to read, create, or change project content.
4. Agent model
Chooses the model endpoint for the conversation.
Use it when: you need to choose the model before the first message; it is locked once the conversation starts.
5. Thinking strength
Sets the model's reasoning effort, affecting response speed and model cost.
Use it when: you want to trade speed and cost for deeper reasoning.
6. Tool permissions
Controls which file, command, network, and remote-device actions Levin may use, and when it must ask you first.
Use it when: you want to review the profile before sending, then adjust it as the task scope or risk changes.
The Agent model is locked after the first message so the conversation keeps a stable runtime. You can still adjust thinking strength whenever Levin is not streaming.
Context and attachment management
The + button adds structured context to the next message. These entries are more precise than explaining every resource in prose, and they remain visible above the editor before you send.
1. Capability
Use @ to add an enabled Skill, Subagent, Workflow, MCP service or tool, or remote device.
2. Structure
Use # to attach a structure loaded in Mol*, either the full structure or the active selection.
3. Attachment
Add an image, text or code file, document, or another supported local file.
4. Write a plan
Start a read-only planning pass to inspect context and prepare a checkpoint before execution.
Plugins are not referenced as a separate attachment type. An enabled plugin contributes Skills, MCP services, or other capabilities; reference the capability that the plugin makes available.
Reference capabilities and structures
Type @ to open the capability picker. Choose one of the five categories below; the selected entry is attached as structured context to the message before you send it.
1. Skills
Provides: Reusable specialist instructions and working methods.
Use it: Select an enabled Skill with @; Levin loads its full instructions when the task matches.
2. Agents
Provides: Focused subagents with their own context and tool preset.
Use it: Delegate a bounded task; the returned result comes back to the main task.
3. Workflows
Provides: Repeatable multi-step procedures with a defined order.
Use it: Choose a Workflow when the same sequence and checkpoints should be reused.
4. MCP
Provides: External services and the tools exposed by each service.
Use it: Select an MCP service, then choose a specific tool before sending.
5. Devices
Provides: Configured remote devices available to the workspace.
Use it: Select a Device when work needs to run through that remote environment.
Mol* structures: after opening a structure, type # to attach the complete structure or the active selection.
Tool permissions
The shield control in Levin's composer is a Sandbox Profile picker. It has three profiles that set the filesystem sandbox, network mode, and approval policy for tool calls. This is separate from Goal mode. New chats use Approved (auto-approve) by default.
Three sandbox profiles
| Profile | Runtime boundary | Approval behavior |
|---|---|---|
Request request-approval | Workspace-write filesystem sandbox with restricted network access. | Asks before local-process internet access, external or unknown tools, expensive compute, and risky commands. Ordinary project-local work can run inside the sandbox. |
Approved auto-approve | Workspace-write filesystem sandbox with network access enabled. | Runs ordinary project-local work automatically and asks only for detected-risk operations, external or unknown tools, high-risk actions, or project-outside writes. |
Full Access full-access | Filesystem sandbox disabled; network access enabled. | Runs without routine tool approval prompts or sandbox restrictions. Host-level safety boundaries can still apply to specific remote operations. |
Note: Changing the profile updates the live session policy for later tool calls; a process that has already started keeps the policy it was launched with. Goal mode manages long-running progress, but it does not grant extra tool access.
Citations and follow-ups
When a line in the conversation deserves another question, you do not need to copy and paste it. Select the text, then choose whether to continue in the current task or open a side chat for a separate follow-up.
-
Follow up in the current task
Select a single line from an assistant reply, reasoning trace, code block, or tool result. In the floating toolbar, choose Add to current task; the passage is added to the composer as a quoted reference beside your draft.
You can select text while a response is streaming too. The visible text is fixed at the moment you click. Reference length and count limits are shown directly when exceeded, and dragging across multiple lines does not open the toolbar.
Figure: the selected passage returns to the current task as context for the next message.
-
Ask about it in a side chat
Select a line in the main conversation and choose Ask in side chat from the floating toolbar. Levin opens an independent side-chat tab and carries the selected passage over as a reference.
The main task keeps running while you ask about a function, tool result, or decision in the side chat. The two conversations stay separate, so the follow-up does not interrupt the main task.
Figure: the selected passage is carried into the side chat for a focused follow-up.
Conversation branching
To try a different direction without losing the current path, hover over a completed assistant response and choose Branch. Levin creates a new task from that response, titles it Fork of <original title>, and marks it as derived from the conversation. The branch copies the conversation history up to that point, along with the model and reasoning settings. Project conversations and their branches continue to use the same project folder; a loose conversation receives a new managed folder. The original conversation remains unchanged and continues normally.
- Only stable, completed assistant responses can be branched. Interrupted or failed partial responses, user messages, and tool results do not offer the action.
- Branching copies conversation history, not disk state. Project conversations and their branches use the same project folder; a loose conversation gets a separate managed folder.
- Queued messages, active work, and background jobs are not copied to the new task.
- Branching is unavailable while context compaction is running.
Manage conversations
A single right-click menu makes it easy to manage the conversation lifecycle.
| Action | Use it for |
|---|---|
| Pin | Keep an active or important conversation at the top of the sidebar. |
| Rename | Give the chat a title that describes its outcome instead of its first message. |
| Assign to project | Move a free chat under a registered project when the work becomes project-specific. |
| Branch | Explore another direction from completed history while leaving the original conversation unchanged. |
| Archive | Remove finished work from the active sidebar without mixing it into current tasks. |
Branching copies conversation state, not disk state. Project chats and their branches continue to use the same project folder unless a separate workspace has been created.
Recommended use
- Start with the outcome. Describe what should be built, fixed, or analyzed. For work that needs several rounds, use Goal mode to define a result Levin can verify.
- Add the relevant context. Use
@to reference capabilities, and attach screenshots, documents, or requirement material when needed. - Use commands and Skills to move faster. Type
/for a command or$to reuse a Skill. - Define the permission boundary. For work that needs ongoing management, set a Goal and check approval permissions before allowing file, command, network, or remote-device actions.
- Keep one task moving continuously. Follow up, add constraints, review changes, and complete pre-submit checks in the same task so the context stays together.