Quickstart
Connect your host to the sandbox, limit it to read tools, and get a first answer from a synthetic user's record
Connect an MCP host to the b.well sandbox, limit it to tools that read data, and check that the model answers from a synthetic user's record. The steps work with any host that supports remote MCP servers over Streamable HTTP.
1. Get sandbox access
Request sandbox access and a synthetic user through Testing Considerations. Sign in to the sandbox only as a synthetic user.
2. Connect your host
Get the synthetic user's access token with the authentication pattern you choose in MCP Authentication. We recommend Federated Authentication if your platform can support it.
- With Federated Authentication, or Static Client Registration if b.well has set it up for you, connect your host to
https://api.client-sandbox.icanbwell.com/mcpand send the access token as a bearer token. - With Dynamic Client Registration, add
https://mcpfhiragent.client-sandbox.icanbwell.com/to your host. It opens the hosted b.well sign-in page, where you sign in as the synthetic user.
The OpenAI, Anthropic, and Google guides have working setups for those platforms.
Your host's tool list shows the tools enabled for your client in the sandbox, which can differ from MCP Tools.
3. Limit the host to read tools
Keep only tools that read data, such as get_patient_summary, get_fhir_record, get_labs, and check_insurance_coverage, until you've set up approvals. The server confirms only some write tools, and your host, or your application if the host has no approval setting, must ask before the rest. Use your host's tool allowlist or per-tool settings to turn the other tools off.
4. Guide the model
If your host accepts your own instructions, such as a system prompt or project instructions, add the base guidance there. b.well's tool descriptions already tell the model what each tool does, so this step is optional for a first test.
5. Ask a question and check the tool call
Ask Give me a summary of my health. The model should call get_patient_summary. In your host's tool call log, the call's input has no person_id, and its result holds the synthetic user's summary as Markdown, one ## heading per section.
If the result's error reads No data found for the given patient ID and resource type., b.well has no summary for this synthetic user; use a synthetic user that has clinical data. For any other error, look up its code in Errors.
Next steps
- Guidelines and workflows: scope tools to your use case, add write tools safely, and guide the model through each workflow.
- MCP Authentication: get tokens from your own application.
Updated about 17 hours ago

