The Neurosnap Agent can help plan research, inspect available tools, analyze files, and run supported computations. A useful request states the scientific objective, input identities, constraints, and the work you authorize.
Start a session
Sign in and open Products → Neurosnap Agent. Upload files using Upload in the Workspace panel, or reference accessible jobs by ID. Specify which file represents each target, ligand, sequence, or dataset.
For example:
Review the structures in job <job-id> for interface quality.
Compare the native confidence metrics and identify uncertain regions.
Do not submit new jobs. Produce a candidate summary with job and file references.
For a new campaign, state the intended modality, target surface, fixed residues, candidate count, and credit budget. Ask for a small pilot before authorizing broad expansion.
Choose guidance and execution scope
The chat interface offers general chat and workflow guidance for structure prediction, binder design, and enzyme engineering. These choices supply task context; they do not make every predicted result reliable or remove the need to specify inputs.
The reasoning selector affects response effort, latency, and usage. Review the Workspace panel for usage and saved files. Do not assume every action presents a separate confirmation: describe explicitly whether the agent may submit jobs or should only analyze existing results.
Understand files and network access
Session files and published outputs can remain available after the execution sandbox closes. Download research artifacts you need to preserve and keep job IDs with the reports.
The network control offers open and restricted modes. Restricted mode limits sandbox access to selected scientific hosts; it does not mean all network activity is disabled. Downloads or redirects outside allowed hosts may fail. Review Agent data handling before supplying confidential information or choosing integrations.
Follow usage and results
Chat model usage, sandbox execution, and hosted jobs are distinct sources of work. The interface reports them separately; a chat allowance should not be interpreted as the total budget for all jobs launched by the agent.
Inspect returned job IDs and progress. An agent’s proposed configuration is not a completed calculation. An analysis should cite actual outputs and preserve uncertainty, rather than presenting an invented score or unrun experiment as a result.
If interrupted, recover the session and existing job IDs before requesting replacement submissions. For external agent clients, see MCP.