Last reviewed on October 26th, 2026
Overview
Neurosnap Agent connects a conversational AI model with computational tools, research data, and a temporary analysis workspace. This document describes the current architecture and data lifecycle to support technical, security, and AI governance reviews. Read it alongside our Privacy Policy, Terms of Use, and Security Measures.
The Agent is optional. Ordinary Neurosnap jobs and analyses do not require an Agent conversation or a third-party MCP client. Using the Agent sends the information needed for the request to OpenAI. Saved Neurosnap conversations remain available after the temporary Agent session closes.
Architecture and data flows
- Your browser connects to Neurosnap. Neurosnap authenticates the user, applies conversation access rules, and stores conversation messages and metadata in MongoDB. Uploaded and saved generated chat files are stored in Cloudflare R2.
- Neurosnap sends a request to OpenAI. The request includes your prompt, relevant conversation context, workflow guidance, and files or Neurosnap data made available for the work. The current Agent uses OpenAI’s Agents API with an OpenAI-hosted sandbox. Prior context may be supplied again when a session restarts.
- The Agent uses tools and a workspace. It can run analysis code in its sandbox, retrieve accessible Neurosnap job data, and request computational jobs through authenticated backend tools or MCP (Model Context Protocol). Tool results returned to the Agent are processed by OpenAI. Job execution uses the infrastructure associated with the selected Neurosnap service; a job does not necessarily run in the Agent’s sandbox.
- Results return to the conversation. Neurosnap saves responses and imports published deliverables into chat file storage. Temporary working files are not all saved as chat attachments. Source job files remain in job storage unless explicitly copied into the conversation.
External database and website requests create an additional data flow. Depending on the request and network settings, search terms or relevant portions of prompts, sequences, files, or results may be sent to those destinations. See network controls below.
Access and sharing
- Personal jobs: private by default. Enabling the job’s Share feature makes it accessible through its sharing link.
- Team jobs: accessible to all current members of the associated team by default, without separately enabling Share.
- Personal Agent conversations: available to their creator.
- Team Agent conversations: their messages, files, and results are available to all current members of the selected team.
Backend tool access is checked using the acting user’s Neurosnap permissions. Content retrieved from a job and included in a team conversation becomes available within that conversation; consider the conversation audience before requesting an analysis of confidential job data. Team membership changes can change who can access team content. Team credits may pay for a personal conversation without making that conversation visible to the team.
Review the selected job and conversation scope before submitting data. See Sharing Your Content for access settings and your responsibilities.
Retention and deletion
Saved conversations, temporary execution sessions, and provider records have separate lifecycles. Closing a sandbox does not delete the conversation stored on Neurosnap.
| Data | Storage and lifecycle |
|---|---|
| Saved conversations | Messages, conversation metadata, and related workflow records are stored in Neurosnap’s database so you can return to your work. Idle-session cleanup does not expire this history. Successful conversation deletion removes its messages and associated conversation records from active Neurosnap storage. |
| Saved chat files | Uploaded files and imported deliverables are stored in Cloudflare R2. They persist independently of the sandbox. Files can be deleted individually; successful conversation deletion removes the conversation’s associated files from active chat storage. |
| Temporary Agent workspace | The current idle-cleanup threshold is ten minutes of inactivity for eligible sessions. Neurosnap normally requests session deletion after that threshold and also closes sessions during certain workflow transitions. Active work, artifact recovery, provider errors, and cleanup retries can delay closure. This is a cleanup trigger, not a guaranteed erasure deadline. |
| Source jobs | Original jobs, inputs, and outputs have their own lifecycle. Deleting an Agent conversation does not delete the jobs it accessed or submitted, or copies saved elsewhere. |
| Database backups | Database backups are encrypted. The current backup configuration schedules backups every twelve hours and cleanup of backups older than thirty days. Deleted conversation data may remain in earlier backups until scheduled cleanup completes; active-storage deletion does not immediately erase backups. |
| Operational records | Usage, billing, diagnostic, and security records may remain after conversation deletion where needed to operate the service, resolve disputes, prevent abuse, or meet legal obligations. This page does not specify one retention period for every record category; contact us for requirements specific to your organization. |
| OpenAI and external recipients | OpenAI may retain application state, files, and abuse-monitoring records under its applicable policies. Independent websites, databases, and MCP clients apply their own practices. Neurosnap conversation deletion does not guarantee immediate erasure of information already received by those services. |
Conversation deletion may be unavailable while work or file operations are active and may need to be retried if provider cleanup fails. Neurosnap requests deletion of the associated OpenAI session as part of conversation cleanup. A provider deletion request does not establish that all provider security, legal, or backup records have been erased.
Providers, training use, and processing location
As described in our Terms of Use, Neurosnap does not use your data or outputs to train models, improve algorithms, or develop new services without your explicit consent.
OpenAI states that API data is not used to train or improve its models by default unless the API customer explicitly opts in. Training use and data retention are separate: OpenAI’s Agents API retains session state and currently does not support Zero Data Retention. Its published data controls also describe abuse-monitoring retention and applicable exceptions. See OpenAI’s API data controls and Agents API documentation for current provider practices.
OpenAI currently documents Agents API data residency support in the United States only. Other Neurosnap infrastructure and external services have their own processing locations. Restricted network access is not a data-residency control. Contact us before relying on a particular region, retention period, or regulated-data processing arrangement.
Our service provider list identifies providers used across Neurosnap. It does not mean every provider receives every Agent request. The recipients depend on the workflow, tools, and infrastructure used.
Network controls and external MCP clients
Restricted sandbox network access
Restricted mode limits outbound sandbox requests to specified hostnames for Neurosnap and scientific resources including UniProt, PubChem, AlphaFold DB, and the EBI host used by ChEMBL. This is a hostname boundary: an allowed shared host may expose resources beyond the named database.
Restricted mode still permits requests to allowed destinations and does not prevent OpenAI from processing the request or Neurosnap tools from accessing authorized data. It does not provide a guarantee that sensitive content cannot leave Neurosnap. With unrestricted sandbox network access, the Agent can contact additional external destinations. A destination may receive relevant query content and apply its own privacy and retention practices.
Using Neurosnap through another MCP client
If you connect Neurosnap MCP to a third-party client, that client can receive tool descriptions, request arguments, and authorized tool results, including job data and files. It may forward this information to its own model provider or other services. The hosted Neurosnap Agent’s sandbox settings and cleanup schedule do not govern that client. Review the client’s permissions, model provider, retention settings, and policies separately.
Scientific reliability and limitations
Neurosnap’s custom biotech harness guides model tool use, analysis, and outputs. Backend permission checks and submission validation complement the model’s instructions. Specialized modes add workflow protocols for structure and complex prediction, binder design, and enzyme design and optimization. General chat does not select a specialized workflow protocol.
The harness encourages inspection of actual tool outputs, traceable job references, recorded methods, and explicit assumptions and limitations. These controls support review and reproducibility; they do not guarantee correct conclusions, identical repeated outputs, or successful experiments. Model behavior, tool versions, stochastic methods, input quality, and external data can affect results.
The Agent can submit real jobs and take supported actions that change platform data or consume credits. Review consequential requests and validate scientific outputs before relying on them. Computational predictions require appropriate independent scientific review and experimental validation. Agent outputs are not a substitute for medical, clinical, regulatory, or safety advice.
Enterprise review and questions
Our Security Incident Response & Breach Notification Policy explains how we handle incidents affecting customer data, including Agent conversations and research files.
For architecture questions, data-handling questionnaires, or requirements involving retention, processing location, or contractual safeguards, contact hello@neurosnap.ai or our support team. Please identify the feature, intended data types, conversation scope, and controls your organization requires.
This overview describes the implementation as of the review date. Features, providers, and settings can change. Consult the linked policies and any applicable agreement for the commitments governing your use of Neurosnap.