SandboxAQ MCP Server

Overview

The Model Context Protocol (MCP) allows SandboxAQ to connect securely with large language models (LLMs) and AI assistants, enabling scientists and engineers to run high-accuracy scientific calculations conversationally and receive structured, traceable scientific results. The SandboxAQ MCP Server exposes SandboxAQ’s proprietary scientific AI models as callable tools within Claude and other MCP-compatible clients. This enables R&D teams to integrate scientific simulation directly into their existing AI-driven workflows without managing compute infrastructure.

What is the SandboxAQ MCP Server?

The SandboxAQ MCP Server is a standardized, managed connection framework that lets AI chat applications securely invoke SandboxAQ’s proprietary scientific tools. Through the MCP, users and agents can:

Ask scientific questions in natural language (e.g., “What is the most stable binding site for N₂ on a Co-Ni surface?”).
Retrieve structured scientific results — binding energies and per-placement convergence status — formatted for immediate use in LLM reasoning.
Run complex multi-step adsorption workflows without writing code or managing GPU infrastructure.
Integrate SandboxAQ model outputs with other data sources, notebooks, or agentic pipelines.

The SandboxAQ MCP Server enables all model access to happens under SandboxAQ’s security, authentication, and audit controls. No model weights or proprietary data are exposed externally.

How It Works

When a user submits an actionable query for the applicable SandboxAQ model through an LLM connected to the SandboxAQ MCP Server (for example, “Calculate the adsorption energy of CO on a copper 111 surface”), the MCP:

Parses the natural language input and maps it to the appropriate tool — aqcat for surface chemistry, or one of the four AQPotency tools for protein–ligand work.
Validates the request parameters and verifies the user’s OAuth session.
Executes the calculation on SandboxAQ’s managed GPU infrastructure using the AQCat Adsorption Spin model.
Returns a structured JSON response — including binding energies, convergence flags, and metadata — to the LLM.
The LLM incorporates the scientific result into its response, answer, or downstream reasoning.

The entire compute lifecycle happens on SandboxAQ infrastructure.

Prerequisites

To access the SandboxAQ MCP Server through a compatible client, the following must be in place:

An active SandboxAQ MCP Server account, created at platform.aisim.sandboxaq.com/signin.
OAuth credentials — an email/password issued during sign-up, or a Google account for social login. Enterprise SSO via your organization’s identity provider is optional and can be configured from the platform console.
A supported MCP client: Claude Desktop, claude.ai, Claude Science, Claude Code, or Antigravity. Any other client that speaks the Model Context Protocol over Streamable HTTP also works — Antigravity is covered below. See Connect your MCP client for step-by-step setup.

Scope. The pricing, limits, and legal terms (Master Customer Agreement, Data Protection Addendum, Privacy Policy) linked from this site cover the Individual self-serve tier. Enterprise terms are negotiated separately — contact mcp-prod@sandboxaq.com to discuss enterprise contracts, SLAs, custom deployments, or volume commitments.

Using the SandboxAQ MCP Server

The SandboxAQ MCP Server can be used by any AI agent or LLM that supports the Model Context Protocol. This enables researchers, developers, and enterprise teams to invoke SandboxAQ’s scientific AI models directly within their existing AI workflows while maintaining security and IP protection.

Example Scenarios

LLM-Driven Research Assistant: Connect your MCP client to the SandboxAQ MCP Server to answer questions like “Which platinum surface facet binds hydrogen most weakly?” without any manual setup.
Internal Copilot Integration: Embed the SandboxAQ MCP Server into another internal agent to enable scientific planning and post-analysis of adsorption workflows.
Third-Party AI Platform: Integrate SandboxAQ model inference into another MCP-based AI platform under your organization’s governance controls.

Setup Overview

Create your account at platform.aisim.sandboxaq.com/signin. You’ll receive your server URL and credentials, and can optionally configure SSO from the platform console. Enterprise customers with negotiated terms: contact mcp-prod@sandboxaq.com instead.

Configure your MCP client

Follow the client-specific instructions below to add your server URL.

Enter your MCP URL

Self-serve accounts use https://mcp.platform.aisim.sandboxaq.com. Enterprise accounts are issued their own host at sign-up — use the URL shown in your platform console.

Authenticate

SandboxAQ MCP authentication is OAuth-based. Sign in with your email and password, or use Google social login. No API keys are issued or accepted. Enterprise customers configuring SSO through their own identity provider — contact mcp-prod@sandboxaq.com.

Verify the connection

Once connected, ask your client: “What tools do you have available from SandboxAQ?” — it should list the available tools and their descriptions.

Run your first query

Use any of the example prompts in the Usage Examples section below.

Claude Desktop UI (Remote MCP)

Navigate to Settings → Connectors → Add Connector and enter your server URL. Select OAuth as the authentication method and complete sign-in via your organization’s SSO provider when prompted.

Claude Code (CLI)

Add the SandboxAQ MCP Server to Claude Code via the claude mcp add CLI command, or by editing your project’s .mcp.json (or ~/.claude.json) directly.

Connecting the server and authenticating your user are two distinct steps. Adding the entry below wires up the transport; you still need to complete an OAuth handshake before Claude Code can call any tools. Skipping the handshake presents as Unauthorized errors or missing tools when you prompt the model.

// .mcp.json (project-scoped) or ~/.claude.json (user-scoped)
{
  "mcpServers": {
    "sandboxaq-mcp": {
      "type": "http",
      "url": "https://mcp.platform.aisim.sandboxaq.com"
    }
  }
}

To authenticate after the server is configured:

Start Claude Code

Start Claude Code in your project directory (or any directory that inherits the user-scoped ~/.claude.json).

Open the MCP menu

Run /mcp inside the Claude Code session, select sandboxaq-mcp, and choose Authenticate — Claude Code presents the option whenever a server requires auth.

Complete the OAuth handshake in the browser

Claude Code opens the SandboxAQ sign-in page. Sign in with your OAuth credentials (email + password, or Google social login). SandboxAQ does not issue or accept static API keys.

Verify

Back in the Claude Code session, re-run /mcp — the sandboxaq-mcp entry should now show as authenticated and list its available tools. As a functional check, prompt: “What tools do you have available from SandboxAQ?” Claude Code should enumerate aqcat and the four AQPotency tools.

Getting Unauthorized or missing tools after connecting? Re-open the MCP menu with /mcp and re-run the Authenticate step. Adding the server entry to .mcp.json or ~/.claude.json starts the transport but does not sign you in — authentication is a separate action inside the Claude Code session. See also Improving tool-invocation reliability if tools are missing after a successful authentication.

Antigravity

In the Antigravity IDE, open the Agent panel, click the ⋯ menu, then MCP Servers → Manage MCP Servers → View raw config to open mcp_config.json. Add the entry below, save, and click Refresh in the Installed MCP Servers list.

Antigravity uses serverUrl for HTTP MCP servers — not url, which is what the Claude clients use. Watch for this if you’re copy-pasting between client configs.

// ~/.gemini/config/mcp_config.json
{
  "mcpServers": {
    "sandboxaq-mcp": {
      "serverUrl": "https://mcp.platform.aisim.sandboxaq.com"
    }
  }
}

Availability in Other MCP Clients

The SandboxAQ MCP Server is compatible with any client that supports the Model Context Protocol over Streamable HTTP transport. Point your client at your server URL and authenticate when prompted.

Available Tools

The SandboxAQ MCP Server exposes five tools across two Large Quantitative Models (LQMs). All five are read-only inference operations that perform no writes, modifications, or external side effects. Full parameter references and schemas are documented on the AQCat Adsorption Spin and AQPotency pages.

Tool	Model	Description
`aqcat`	AQCat	Fully automated dense adsorption energy screening. Returns the lowest binding energy across placements — with per-placement convergence status — for a given bulk material, adsorbate, and crystal facet.
`AqpotencyPotency`	AQPotency	Predict pIC₅₀ for a single protein–ligand pair from a UniProt accession and a SMILES string.
`AqpotencyScreen`	AQPotency	Score an uploaded SMILES library (up to 10,000 compounds) against a single target for virtual screening.
`AqpotencyScan`	AQPotency	Screen one ligand against a curated target panel (`human_kinome`, `bowes_safety_panel`, or `proteome`) for reverse screening and off-target triage.
`AqpotencySelectivity`	AQPotency	Compute on-target vs. off-target potency and fold-selectivity for one ligand against a named off-target list.

The four AQPotency tools are asynchronous — each returns a job_id in 1–2 seconds, and results are retrieved by polling check_job_status and then calling get_job_results. Your MCP client handles this polling loop for you.

Usage Examples

The following example shows a realistic AQCat prompt you can use once the SandboxAQ MCP Server is connected, along with the expected output. For the full catalog of tool-triggering example prompts, see the AQCat → Example Prompts section and the AQPotency → Example Use Cases section.

Getting your client to call the tool consistently? LLM tool selection is non-deterministic — see Improving tool-invocation reliability below for the fix (enable tool discovery on your client, plus prompt-phrasing tips).

1. Find the most stable N₂ adsorption site on a Co-Ni alloy

User prompt

Use the AQCat tool to find the lowest-binding-energy configuration for N₂ adsorbed on a cobalt-nickel (111) surface. Report the binding energy of the winning placement and confirm whether the relaxation converged.

What your client doesCalls aqcat with bulk_composition_contains: "Co-Ni", adsorbates: ["*N2"], facets: [[1,1,1]]. Evaluates 5 placements and returns the global minimum.Expected output

AQCat evaluated 5 placements for N₂ on Co-Ni (111) (bulk reference `mp-1183837`); the lowest-energy configuration is placement 4 with a **binding energy of approximately −0.43 eV**. All returned placements relaxed to convergence within the step budget (`is_relaxed: true`).

Limits & Quotas

The following limits apply to all requests on the MCP endpoint. These are enforced server-side.

Limit	Value
Bulks per request	1
Adsorbates per request	1
Facets per request	1
Placements per call	5 (hard cap)
Max optimization steps	100 per placement / 500 total
Target latency	<15s per call / <2 min total
Max response tokens	25,000

Pricing

Each model is metered on its own unit. There is no platform fee on the Individual tier.

Model	Billing unit	Rate
AQCat (`aqcat`)	Per Relaxation Unit (RU) — 1 bulk × 1 facet × 1 adsorbate, ≤500 total steps	$1.00 / RU
AQPotency (all four tools)	Per protein–ligand pair evaluated, tiered per call	$0.01 / pair for the first 10 pairs of a call, then $0.001 / pair

Full rate card, worked billing scenarios, and the Individual vs. Enterprise tier comparison are on the Pricing page. Detailed usage logs (timestamp, organization, molecule count, compute duration) are available upon request for billing reconciliation. For enterprise accounts and volume pricing, contact mcp-prod@sandboxaq.com.

Troubleshooting

Troubleshoot connection issues

If you’re having trouble connecting to a service, try these steps:

Check that you have a stable internet connection.
Verify you have the right SandboxAQ-provided MCP URL.
Review any permissions or account type requirements for the service.
If authentication fails, try disconnecting and reconnecting from Settings → Connectors → Add Connector.

Desktop App Showing ‘Disconnected’ After Successful Web Authentication

After disconnecting and reconnecting the SandboxAQ MCP Server from Claude Desktop, the Anthropic authorization popup may confirm the connection succeeded while the Claude Desktop app continues to display the connector as not connected. This is a known Claude Desktop UI-refresh delay — the underlying OAuth handshake completed successfully; the desktop client just failed to pick up the new state.

Quick fix: toggle the connector off and back on in Claude Desktop settings. In Settings → Connectors → SandboxAQ MCP Server, toggle the connector off, wait a moment, then toggle it back on. This forces the desktop client to re-read its connector state and pick up the successful OAuth handshake.

If the toggle doesn’t clear it, sign out of Claude Desktop and sign back in. Fully signing out and back in forces the desktop client to re-fetch its connector state and the active tool list, at which point the SandboxAQ MCP Server will show as connected. Simply quitting and reopening the app is not always sufficient. If the connector still shows as disconnected after signing back in, remove the connector entirely from Settings → Connectors, then re-add it via Add Connector using your SandboxAQ-provided MCP URL.

Improving tool-invocation reliability

Large language models are non-deterministic when deciding whether to invoke an MCP tool. Even when a prompt clearly maps to a SandboxAQ MCP tool (aqcat, AqpotencyPotency, AqpotencyScreen, AqpotencyScan, AqpotencySelectivity), the model may occasionally answer from its own knowledge instead of calling the tool. The recommendations below make tool invocation substantially more consistent. The setup differs by client: Claude Desktop exposes a UI toggle; Claude Code requires an environment variable.

Claude Desktop — disable “lazy loading” of MCP tools. Lazy loading defers tool-schema loading until the model believes a tool is needed — which itself relies on the same non-deterministic decision. Turn it off so all SandboxAQ MCP tool schemas are always loaded and available for selection. In Settings → Connectors → SandboxAQ MCP Server, toggle Lazy load tools to off, then restart Claude Desktop.

Claude Code — set ENABLE_TOOL_SEARCH=true. Claude Code has no “lazy load tools” UI toggle; tool discovery for this MCP server is controlled by an environment variable. Without it, Claude Code may fail to fetch the SandboxAQ MCP tool schemas and can surface as tool-search errors or intermittent 429 rate-limiting during initialization. Export the variable in your shell before launching Claude Code, or add it to the server’s env block in your Claude Code MCP config.

# Shell (bash / zsh) — export before launching Claude Code
export ENABLE_TOOL_SEARCH=true

# Or add to the server entry in .mcp.json / ~/.claude.json:
#   "sandboxaq-mcp": {
#     "type": "http",
#     "url": "https://mcp.platform.aisim.sandboxaq.com",
#     "env": { "ENABLE_TOOL_SEARCH": "true" }
#   }

Prompt-phrasing tips that reliably trigger the correct tool:

Name the tool explicitly. Prefer “Use the AQCat tool to calculate the binding energy of N₂ on Co-Ni (111)” over “What is the binding energy of N₂ on Co-Ni (111)?”
Include the required parameters up front. For AQCat: bulk composition, adsorbate, and facet. For AQPotency: SMILES and UniProt ID (or panel name).
Ask a scoped, computational question rather than an open-ended one. “Screen these 500 SMILES against P00533 with AqpotencyScreen” triggers the tool more reliably than “Which of these compounds bind EGFR?”
Point at the SandboxAQ MCP Server by name when a session has multiple MCP servers connected — the model otherwise weighs across all available tools.
If the model still doesn’t call the tool, explicitly ask it to. “Please call the aqcat tool with the parameters above” is a valid recovery prompt.

If the issue persists, contact support.aisim@sandboxaq.com.

Permissions & Security

MCP queries run on behalf of the authenticated user, verified via OAuth 2.0. All compute runs on SandboxAQ-managed infrastructure — model weights are not transmitted to or accessible by end users.

IP Protection: Model weights are encrypted and not exposed via the MCP interface. All inference occurs within SandboxAQ’s secure compute environment.
OAuth 2.0: Authentication is handled through OAuth — email/password, Google social login, or enterprise SSO via your organization’s identity provider. No API keys are issued or stored.
Organizational Isolation: Each organization’s requests and usage data are isolated to prevent cross-organization data access.
Enterprise VPC deployment: High-security enterprise clients may request deployment of the AQCat inference stack within their own VPC, providing zero-data-egress capabilities. This is an enterprise-only offering — contact mcp-prod@sandboxaq.com to scope a customer-VPC deployment.

Privacy & Data

See the SandboxAQ Privacy Policy and the Data Protection Addendum for full details on how SandboxAQ collects, uses, and protects data related to your use of the MCP Server.

Get Started

MCP Server

Models

Support & Legal

Overview

What is the SandboxAQ MCP Server?

How It Works

Prerequisites

Using the SandboxAQ MCP Server

Example Scenarios

Setup Overview

Claude Desktop UI (Remote MCP)

Claude Code (CLI)

Antigravity

Availability in Other MCP Clients

Available Tools

Usage Examples

Limits & Quotas

Pricing

Troubleshooting

Troubleshoot connection issues

Desktop App Showing ‘Disconnected’ After Successful Web Authentication

Improving tool-invocation reliability

Permissions & Security

Privacy & Data

​Overview

​What is the SandboxAQ MCP Server?

​How It Works

​Prerequisites

​Using the SandboxAQ MCP Server

​Example Scenarios

​Setup Overview

​Claude Desktop UI (Remote MCP)

​Claude Code (CLI)

​Antigravity

​Availability in Other MCP Clients

​Available Tools

​Usage Examples

​Limits & Quotas

​Pricing

​Troubleshooting

​Troubleshoot connection issues

​Desktop App Showing ‘Disconnected’ After Successful Web Authentication

​Improving tool-invocation reliability

​Permissions & Security

​Privacy & Data

Overview

What is the SandboxAQ MCP Server?

How It Works

Prerequisites

Using the SandboxAQ MCP Server

Example Scenarios

Setup Overview

Claude Desktop UI (Remote MCP)

Claude Code (CLI)

Antigravity

Availability in Other MCP Clients

Available Tools

Usage Examples

Limits & Quotas

Pricing

Troubleshooting

Troubleshoot connection issues

Desktop App Showing ‘Disconnected’ After Successful Web Authentication

Improving tool-invocation reliability

Permissions & Security

Privacy & Data