> ## Documentation Index > Fetch the complete documentation index at: https://docs.aisim.sandboxaq.com/llms.txt > Use this file to discover all available pages before exploring further. # Connect your MCP client > Step-by-step guides for connecting the SandboxAQ MCP Server to Claude Desktop, claude.ai, Claude Science, Claude Code, and Antigravity. ## Before you connect This guide shows how to connect the SandboxAQ MCP Server to the client you work in. The SandboxAQ OAuth flow is the same everywhere — only the interface around it differs, so pick your client below and follow that section. **In a hurry?** The [Quickstart](/getting-started/quickstart) has the condensed version — a handful of steps per client and nothing else. This page adds screenshots, client-specific caveats, and troubleshooting. Client interfaces change frequently, so treat the steps below as a general guide. For the most accurate instructions, check the documentation for the client you're using. ### Sign up and get your MCP URL Every client needs two things: your server URL, and an account to authenticate against. Both come from sign-up. Go to [platform.aisim.sandboxaq.com/signin](https://platform.aisim.sandboxaq.com/signin) and sign up. You can register with an email address and password, or use Google social login. Review and accept the terms on the subscription page to activate your account. A card is required at sign-up, but nothing is charged against it while your free credits last. Once the credits are exhausted, usage is billed pay-and-deduct and requests fail with a `402` error if the balance reaches zero. See [Pricing](/getting-started/pricing) for rates. In the platform console, use the **Copy MCP URL** button in the top-right corner, or open **Account → MCP**. Keep it on your clipboard — every client below asks for it. Self-serve accounts all share the same server URL — `https://mcp.platform.aisim.sandboxaq.com` — which is what this page uses throughout. Enterprise accounts are issued their own host; use the URL shown in your platform console instead. **Scope.** The pricing, limits, and legal terms ([Master Customer Agreement](/support-legal/mca), [Data Protection Addendum](/support-legal/dpa), [Privacy Policy](/support-legal/privacy)) linked from this site cover the **Individual self-serve** tier. **Enterprise** terms are negotiated separately — contact [mcp-prod@sandboxaq.com](mailto:mcp-prod@sandboxaq.com) to discuss enterprise contracts, SLAs, custom deployments, or volume commitments. ### How authentication works Many scientific APIs authenticate with a static key passed as a request header. **The SandboxAQ MCP Server does not** — no API keys are issued, stored, or accepted. Every client authenticates *you* over **OAuth 2.0**: email and password, Google social login, or enterprise SSO through your organization's identity provider. Practically, this means there is no secret to copy into a config file. The only thing you paste anywhere is your MCP URL; the sign-in happens in a browser window that your client opens for you, and the resulting session is scoped to your user and your account. **Connecting the server and authenticating your user are two distinct steps.** Adding the server wires up the transport; you still need to complete an OAuth handshake before the client can call any tools. Skipping the handshake presents as `Unauthorized` errors or missing tools when you prompt the model. | Client | Where you authenticate | What "connected" looks like | | ---------------------------------------------- | ------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------- | | Claude Desktop / [claude.ai](http://claude.ai) | A browser tab opened by **Add connector** — click **Allow Access**, then sign in | The connector shows **Connected ✓**, and its toggle is blue under **+ → Connectors** in a chat | | Claude Science | Inherited from your Claude account — the app asks to **Authorize** access to your connectors on first launch | **SandboxAQ** is listed and toggled on under **Connectors & skills** | | Claude Code | `/mcp` → select `sandboxaq-mcp` → **Authenticate** → browser sign-in | `/mcp` lists the server as authenticated and shows its tools | | Antigravity | A browser prompt after you add the server and click **Refresh** | The server appears under **Installed MCP Servers** with its tools listed | ## Set up your client The SandboxAQ OAuth flow is identical in every client; only the surrounding interface changes. Pick the client you work in. Use this method to connect from the standard Claude chat interface. The SandboxAQ MCP Server is a **remote connector**, and connectors are stored on your Claude account — so the steps below are identical in the Claude Desktop app and at [claude.ai](https://claude.ai), and configuring one configures both. Open [claude.ai](https://claude.ai) or the Claude Desktop app and sign in with your Anthropic account. Click your profile icon, choose **Settings**, then go to **Customize → Connectors** — or go directly to [claude.ai/customize/connectors](https://claude.ai/customize/connectors). Click **Add → Browse connectors**, search for **SandboxAQ**, and select it. Click **Connect**, paste the MCP URL you copied above, and click **Continue**. If your workspace administrator has already published the connector, it appears in the list directly — just click **Connect**. A new browser tab opens. Click **Allow Access**, then sign in with your SandboxAQ credentials (email and password, or Google). When the handshake finishes you are returned to Claude and the connector shows **Connected ✓**. Open a new chat, click the **+** button at the bottom of the composer, choose **Connectors**, and confirm the SandboxAQ toggle is on (blue). Claude now has access to the SandboxAQ tools in that conversation. **Tired of approving every call?** The first time Claude invokes a SandboxAQ tool it asks you to confirm, and a single prompt often triggers several tool calls. To stop re-approving, go back to **Settings → Connectors**, select the SandboxAQ connector, and switch **Needs approval** to **Always Allow**. **Turn off "Lazy load tools."** Lazy loading defers loading the tool schemas until Claude thinks it needs them, which makes tool selection less reliable. In **Settings → Connectors → SandboxAQ MCP Server**, set **Lazy load tools** to **off** and restart Claude. See [Improving tool-invocation reliability](/mcp-server/mcp-server#improving-tool-invocation-reliability) for the full explanation. **Verify the connection.** Ask Claude: ```text theme={null} What tools do you have available from SandboxAQ? ``` Claude should enumerate `aqcat` and the four AQPotency tools — `AqpotencyPotency`, `AqpotencyScreen`, `AqpotencyScan`, and `AqpotencySelectivity` — along with their descriptions. **Try it:** *"Predict the binding potency of CC1=CC=C(C=C1)S(=O)(=O)N against protein P00533."* More in [Run your first query](#run-your-first-query). [Claude Science](https://claude.com/product/claude-science) is a desktop research application that runs analyses, searches databases, and traces its work. It uses the connectors on your Claude account rather than defining its own, so the SandboxAQ MCP Server carries over automatically once it is connected. **Prerequisite.** Complete the **Claude Desktop and claude.ai** tab first. Claude Science has no place to paste an MCP URL — it reads the connectors already attached to your Claude account. Go to [https://claude.com/product/claude-science](https://claude.com/product/claude-science), click **Download now**, and install the app for your platform. Launch Claude Science and choose **Sign in with Claude.ai** — use the same Claude account that holds the SandboxAQ connector. (**Paste code instead** is the fallback if the app cannot open a browser for you.) In the browser, Claude Science asks to connect to your Claude chat account and lists the permissions it needs — including **Use and manage your connectors**, which is what carries the SandboxAQ connection across. Click **Authorize**. On **Set up Claude for your science**, click **Continue**. On **Connect to the scientific web**, choose which public sources Claude may reach — papers, sequences, structures, and the databases behind them — then click **Continue**. On the **Connectors & skills** screen, find **SandboxAQ** in the connector list and make sure its toggle is on, then click **Continue**. Describe your field on **What do you work on?** (for example, *"I am a catalysis scientist working with homogeneous and heterogeneous systems"*). On **Where should we start?**, pick a first task — **Run in silico experiments with MCP server tools** is the relevant one here — and click **Start**. **Turn off "Lazy load tools."** Claude Science inherits this setting along with the connector, so it has to be changed on your Claude account — in **Settings → Connectors → SandboxAQ MCP Server** at [claude.ai](https://claude.ai) or in Claude Desktop, not in Claude Science itself. See [Improving tool-invocation reliability](/mcp-server/mcp-server#improving-tool-invocation-reliability) for the full explanation. **Verify the connection.** In a Claude Science session, ask: ```text theme={null} What tools do you have available from SandboxAQ? ``` Claude should list `aqcat` and the four AQPotency tools. If it does not, open Claude Science's settings and re-check the connector toggle under **Connectors & skills**, confirm the connector still shows **Connected ✓** at [claude.ai/customize/connectors](https://claude.ai/customize/connectors), and check that **Lazy load tools** is off. **Try it:** *"Scan CC1=CC=C(C=C1)S(=O)(=O)N against the `human_kinome`."* More in [Run your first query](#run-your-first-query). Claude Code connects to the SandboxAQ MCP Server over Streamable HTTP. The only prerequisite is a working Claude Code installation. **Register the server** (enterprise accounts: substitute your own host in the URL): ```bash theme={null} claude mcp add sandboxaq-mcp --transport http https://mcp.platform.aisim.sandboxaq.com # Confirm it was registered claude mcp list ``` **Scope.** Claude Code asks where to register the server — local to the current directory, shared with the project, or for your user account. Choose **user** scope if you want the server available everywhere. Restart Claude Code after the change so it picks up the new configuration. If you prefer editing configuration directly, the equivalent entry is: ```json theme={null} // .mcp.json (project-scoped) or ~/.claude.json (user-scoped) { "mcpServers": { "sandboxaq-mcp": { "type": "http", "url": "https://mcp.platform.aisim.sandboxaq.com" } } } ``` Registering the server is not the same as signing in — complete the handshake before prompting: Start Claude Code in your project directory (or any directory that inherits the user-scoped `~/.claude.json`). Run `/mcp` inside the session to open the MCP server menu. Select `sandboxaq-mcp` and choose **Authenticate** — Claude Code offers the option whenever a server requires auth. Claude Code opens the SandboxAQ sign-in page. Click **Allow Access** and sign in with your SandboxAQ credentials (email and password, or Google). No API key is involved. Back in the session, re-run `/mcp` — `sandboxaq-mcp` should now show as authenticated and list its tools. **Set `ENABLE_TOOL_SEARCH=true` before launching Claude Code.** Claude Code has no "lazy load tools" UI toggle; tool discovery is controlled by this environment variable, and without it Claude Code may fail to fetch the SandboxAQ tool schemas. See [Improving tool-invocation reliability](/mcp-server/mcp-server#improving-tool-invocation-reliability) for the shell export and the `env` block variant. **Verify the connection.** As a functional check, prompt: ```text theme={null} What tools do you have available from SandboxAQ? ``` Claude Code should enumerate `aqcat` and the four AQPotency tools — `AqpotencyPotency`, `AqpotencyScreen`, `AqpotencyScan`, and `AqpotencySelectivity`. **Try it:** *"Use the AQCat tool to find the lowest-binding-energy configuration for N₂ on a cobalt-nickel (111) surface."* More in [Run your first query](#run-your-first-query). Antigravity connects to remote MCP servers through a raw JSON config file rather than a connector UI, and it uses the `serverUrl` key where the Claude clients use `url`. Open the **Agent** panel, click the **⋯** menu, then choose **MCP Servers → Manage MCP Servers → View raw config**. This opens `~/.gemini/config/mcp_config.json`. Add the entry below and save the file. ```json theme={null} // ~/.gemini/config/mcp_config.json { "mcpServers": { "sandboxaq-mcp": { "serverUrl": "https://mcp.platform.aisim.sandboxaq.com" } } } ``` Back in **Manage MCP Servers**, click **Refresh** in the Installed MCP Servers list. `sandboxaq-mcp` should appear. Complete the OAuth prompt when it appears and sign in with your SandboxAQ credentials (email and password, or Google). No API key is involved. **`serverUrl`, not `url`.** Antigravity rejects the `url` key that the Claude clients use. This is the most common reason a server never appears in the list. **Two limitations to know before you start.** File upload is not supported, so `AqpotencyScreen` cannot run at all — use [`AqpotencyScan`](/models/aqpotency#tools) or run the screen from a Claude client instead. Tool invocation is also inconsistent, so name the tool explicitly and supply every required parameter up front. Details: [Platform Compatibility](/mcp-server/platform-compatibility). **Verify the connection.** Prompt: ```text theme={null} What tools do you have available from SandboxAQ? ``` **Try it:** *"Scan CC1=CC=C(C=C1)S(=O)(=O)N against the `human_kinome`."* More in [Run your first query](#run-your-first-query). ## Run your first query Once the tools are listed, you can run real calculations by describing what you want in natural language. The prompts below work identically in Claude Desktop, Claude Science, and Claude Code. Naming the tool explicitly — as these do — makes invocation substantially more reliable. Each prompt below shows what comes back and what it costs. Run all four and you'll have spent under **\$2** in total. ### AQPotency — protein–ligand potency, panels, and selectivity Predict the potency of a single compound against a single protein target: ```text theme={null} Predict the binding potency of CC1=CC=C(C=C1)S(=O)(=O)N against protein P00533 ``` **You get back** a predicted pIC₅₀ for that pair, its uncertainty (`potency_sigma`), and two applicability flags — `ligand_similarity` and `protein_in_training` — that tell you how much to trust the number. Cost: **\$0.01**. Scan one compound across a whole protein panel: ```text theme={null} Scan CC1=CC=C(C=C1)S(=O)(=O)N against the human_kinome ``` **You get back** all 604 kinases in the panel scored and ranked by predicted potency, so you can see what else the compound is likely to hit. Cost: **\$0.694**. Assess selectivity against a named set of off-targets: ```text theme={null} Assess the selectivity of CC1=CC=C(C=C1)S(=O)(=O)N against target Q15078 vs. off-targets Q6J9G0, Q13546, Q15768 ``` **You get back** the on-target potency, each off-target's potency, and the fold-selectivity between them — including which off-target is the worst offender. Cost: **\$0.04**. ### AQCat — adsorption energies on catalyst surfaces ```text theme={null} 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. ``` **You get back** the lowest-energy placement out of the five AQCat evaluates, its binding energy in eV, and a per-placement convergence flag so you can tell whether each relaxation finished within the step budget. Cost: **\$1.00** (one relaxation unit). For the full catalog of tool-triggering prompts and worked examples with expected output, see [AQCat → Example Prompts](/models/aqcat#example-prompts) and [AQPotency → Example Use Cases](/models/aqpotency#example-use-cases). ## Other MCP clients The SandboxAQ MCP Server works with any client that speaks the Model Context Protocol over Streamable HTTP transport. Point the client at your server URL and authenticate when prompted. * **Anything else** — use your client's remote-MCP or HTTP-transport configuration with the URL `https://mcp.platform.aisim.sandboxaq.com`, then complete the OAuth prompt when it appears. Config key names differ between clients: the Claude clients use `url`, Antigravity uses `serverUrl`. Watch for this if you're copy-pasting between configs. ## If something doesn't work * **The connector shows "disconnected" even though the browser said it worked.** A known Claude Desktop UI-refresh delay. Toggle the connector off and back on in **Settings → Connectors**; if that doesn't clear it, sign out of Claude Desktop and back in. Full steps: [Desktop app showing 'Disconnected'](/mcp-server/mcp-server#desktop-app-showing-disconnected-after-successful-web-authentication). * **`Unauthorized`, or no tools listed, in Claude Code.** Registering the server starts the transport but does not sign you in. Re-open `/mcp`, select `sandboxaq-mcp`, and run **Authenticate** again. * **Claude answers from its own knowledge instead of calling a tool.** Turn off **Lazy load tools** (Claude Desktop and Claude Science) or set `ENABLE_TOOL_SEARCH=true` (Claude Code), name the tool explicitly in your prompt, and include the required parameters up front. See [Improving tool-invocation reliability](/mcp-server/mcp-server#improving-tool-invocation-reliability). * **A call fails with `402`.** Your balance is exhausted. Add or update your payment method in the platform console; see [Pricing](/getting-started/pricing). * **You pasted the wrong URL.** Re-copy it from **Account → MCP** in the platform console. It always has the form `https://mcp.platform.aisim.sandboxaq.com`. Still stuck? Contact [support.aisim@sandboxaq.com](mailto:support.aisim@sandboxaq.com), or use the [contact form](/support-legal/contact). ## Next steps How the server works, the full tool list, limits and quotas, security model, and support terms. Parameter reference, supported bulks and adsorbates, and worked adsorption examples. Potency, screening, panel scan, and selectivity tools, with panel definitions and output schemas. Per-unit rates, tiers, and sample billing scenarios. Short answers on supported clients, account setup, accuracy, privacy, and billing. Enterprise sales, technical support, and privacy inquiries.