Lévin™ Harness, at a glance
Built for open scientific components
Bring the open models, scientific tools, and domain methods you trust into one connected workspace.
Models reason. Tools act. Methods set the boundaries. Lévin™ Harness brings them into one connected workflow, linking conversations, project files, permission checks, and results. You set the goal and make the final decisions. Levin selects the capabilities that fit the context and keeps each step clear.
A RESEARCH LOOP
From scientific context to reproducible execution
Levin organizes research objects, specialist tools, and compute resources in one task, turning a single conversation into an executable, reviewable scientific workflow.
-
1
Establish scientific context
Connect project files, molecular structures, selections, and measurements so Levin can reason about the object under study.
-
2
Compose research capabilities
Enable the model, Skills, Agents, MCP services, workflows, and scientific software needed for the task.
-
3
Execute and preserve the trail
Run on this Mac or a remote GPU, confirm sensitive permissions, and keep results, state, and provenance ready for review.
Chat workspace
Every conversation has its own project scope. Start a new chat and Levin prepares a managed project. When you need to work with existing code, data, or structure files, choose the folder that contains them. The conversation and its tool calls stay within that scope.
The file panel lets you search, preview, edit, save, rename, copy, move, and delete files, or open them in another app. Reference a file path in chat and Levin can read it, make changes, and run checks within the current project and permission boundaries. There is no need to paste the entire file into the prompt.
Molecular context
The Mol* workspace shares structural context with the current conversation. Open or drop in PDB, mmCIF, GRO, XYZ, MOL, SDF, and MOL2 files; fetch a structure by PDB ID; or ask the Agent to load one directly. Structures stay with the current conversation and project scope.
| Surface | What it supports | Agent context |
|---|---|---|
| Structure | Open a local file, drop in a structure, fetch by PDB ID, or load one through the Agent. | Type # to reference the full loaded structure or the current selection. |
| Sequence | View chain sequences, click or drag to select residues, and inspect multiple-structure alignments. | The selected chain and residue range can be included with the message. |
| Measurements | Create labels, distances, angles, dihedrals, orientations, and best-fit planes. | Tool results can retain the measurement type, value, and structural location. |
| Representations | Adjust cartoon, ball-and-stick, surface, coloring, selection representations, and camera views. | Tools can read representation state; vision-capable models can also inspect the rendered view. |
Mol* does not flatten research state into a screenshot. Loaded structures, active selections, and measurements can be attached as structured information, so the Agent knows the exact chain, residues, and measurement you mean.
Reusable workflows
Workflows put familiar methods on a visual canvas. Each card can call an MCP tool, delegate to a Skill, or hold a manual instruction. Connections define order and conditional branches. When you build the workflow, Levin compiles the graph into an enabled Skill.
| Stage | What happens | Result |
|---|---|---|
| Design | Add MCP, Skill, or manual steps, then connect their order and branches. | A process graph you can inspect directly. |
| Build | Validate step and tool references, then compile the graph into a Skill. | An enabled SKILL.md that can be called from chat. |
| Test | Have the active Agent execute every step with sample input and report each result. | The outcome of every step, including failures that need attention. |
| Run | The Agent reads context, completes tool parameters, and follows conditional branches. | It pauses to ask whenever input or confirmation is required. |
Workflows are currently interpreted and run by the Agent; they are not a separate deterministic runtime. Select one explicitly with @, or let the Agent choose an enabled workflow when its description matches the task. Every tool call still follows the active permission boundary.
Integrate open scientific tools
Bring validated open-source scientific software into Levin. Plugins provide executable tools and runtimes. Skills provide domain methods. MCP connects local or remote services. Workflows bring those capabilities together. Results return to the current conversation, and file outputs can stay in the project workspace.
| Integration | Best for | How it enters the workflow |
|---|---|---|
| Plugin SDK | Python packages, command-line programs, models, and long-running scientific jobs. | Declare tools, runtimes, job state, errors, and artifacts for the Agent to use. |
| Skills | Domain guidance, input constraints, analytical methods, and standard operating procedures. | Enable a Skill to add it to the Agent's context, or select it explicitly with @. |
| MCP services | Local processes, remote MCP services, scientific databases, and experimental platforms. | Connect the service and choose a tool. Use @ to select the service or a specific tool. |
| Workflows | Processes that combine tools, branch on conditions, or pause for human confirmation. | Arrange, build, and test the process on the canvas, then let the Agent run it step by step. |
| Runtime devices | This Mac, plus SSH-connected hosts or GPUs configured for plugins. | Select a remote device with @. Deployment and access still follow the configured permission boundary. |
Can't find the tool you need? Share its GitHub repository with the built-in levin-plugin-creator. It reads the actual entry points, dependencies, device requirements, and outputs, then confirms the tool interface, deployment target, and required permissions with you. Only then does Levin generate the plugin, check its interface, package it, deploy it, and verify a real run. Once verified, the tool is ready to use in Chat or a Workflow.
