Probe works with any AI assistant that can connect to a remote MCP server (Model Context Protocol) over HTTPS and send an access key in a custom header. This page covers the assistants we see most often. For Claude Code, see Connect Claude Code to Probe.
What we have tested. Claude Code and the OpenAI Codex CLI both work with Probe. GitHub Copilot CLI is in use with Probe at a customer site. The other setups below follow each assistant's own documentation and have not been tested yet. An assistant that can only sign in with OAuth, and cannot send a custom header, won't work. If yours doesn't connect, contact us and use Claude Code in the meantime.
Step 1: Get your connection details
On your welcome page (https://<your-company>.trial.esharp.se/), press Create my key. Copy the key right away: it is shown only once. If you lose it, press Create a new key; the old key stops working at once. Besides the Claude Code commands, the page shows a ready-made .mcp.json block with both servers and your key. Assistants that read the mcpServers format can use it as it is.
You connect two servers, with the same key for both:
|
Server |
Address |
What it does |
|---|---|---|
|
probe |
|
The board data, review findings and datasheets. |
|
trace |
|
Handles files in the background. Probe needs it, so add both. |
The key is sent in the header X-Esharp-Access, with the key as its value and nothing in front of it. Don't send it as Authorization: Bearer. In the examples below, replace <your-company> with your company's name from the welcome page address, and <your-key> with your key.
Your key is personal, like a password. Keep it out of shared files and code repositories. Where an assistant can read the key from an environment variable, as Codex can, use that.
Probe offers a few hundred tools, and its tool list is large. On a slow connection or a VPN, loading it can take 20 seconds or more, so allow a generous start-up timeout where the assistant has one.
Step 2: Add Probe to your assistant
OpenAI Codex CLI (tested)
Store the key in an environment variable once. On Windows:
setx ESHARP_ACCESS "<your-key>"
On macOS or Linux, export ESHARP_ACCESS from your shell profile instead. Then add to ~/.codex/config.toml:
[mcp_servers.probe]
url = "https://probe.<your-company>.trial.esharp.se/mcp"
env_http_headers = { "X-Esharp-Access" = "ESHARP_ACCESS" }
startup_timeout_sec = 120
[mcp_servers.trace]
url = "https://trace.<your-company>.trial.esharp.se/mcp"
env_http_headers = { "X-Esharp-Access" = "ESHARP_ACCESS" }
startup_timeout_sec = 120
env_http_headers maps the header to the name of the variable, so the key itself never sits in the file. codex mcp add --url can't add a custom header, so edit the file instead.
-
After setting or changing the variable, stop Codex completely. Codex keeps a background process that holds the environment it started with. On Windows:
Get-Process codex | Stop-Process. Then start Codex again. -
If Probe shows only four tools, Codex is connecting without your key: the variable isn't set in the process Codex runs in. Stop Codex as above.
-
If Trace or Probe shows "connected (0 tools)", the tool list timed out. Check that
startup_timeout_sec = 120is in place.
GitHub Copilot CLI
Add to ~/.copilot/mcp-config.json. The top-level key is mcpServers:
{
"mcpServers": {
"probe": {
"type": "http",
"url": "https://probe.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
},
"trace": {
"type": "http",
"url": "https://trace.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
}
}
}
If your company manages Copilot centrally. On Copilot Business or Enterprise, your administrator must allow MCP servers. If your company also runs an MCP registry, Copilot keeps only servers whose name in this file matches the name in the registry. Under any other name, the server connects, lists its tools and is dropped about two seconds later, which looks like a broken server. Ask your administrator for the exact registered names and use them in place of probe and trace.
GitHub Copilot in VS Code (untested)
Run MCP: Open User Configuration from the Command Palette and add. Note that the top-level key here is servers, not mcpServers:
{
"servers": {
"probe": {
"type": "http",
"url": "https://probe.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
},
"trace": {
"type": "http",
"url": "https://trace.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
}
}
}
Use Copilot Chat in Agent mode. The same administrator setting and registry naming as for Copilot CLI apply. Copilot Chat limits how many tools it uses at once, and Probe offers more than that; if Copilot asks you to choose, keep Probe's tools selected and switch off tools you don't need.
Cursor (untested)
Add to ~/.cursor/mcp.json (the global file, so the key never ends up in a project), then restart Cursor:
{
"mcpServers": {
"probe": {
"url": "https://probe.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
},
"trace": {
"url": "https://trace.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
}
}
}
Google Gemini CLI (untested)
Run once:
gemini mcp add --transport http --header "X-Esharp-Access: <your-key>" probe https://probe.<your-company>.trial.esharp.se/mcp
gemini mcp add --transport http --header "X-Esharp-Access: <your-key>" trace https://trace.<your-company>.trial.esharp.se/mcp
A local model in LM Studio (untested)
If design data may not be sent to a cloud model, you can run a model on your own computer. LM Studio 0.3.17 or later can connect to Probe. Open Program › Install › Edit mcp.json and add under mcpServers:
"probe": {
"url": "https://probe.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
},
"trace": {
"url": "https://trace.<your-company>.trial.esharp.se/mcp",
"headers": { "X-Esharp-Access": "<your-key>" }
}
Design review takes many tool calls per question, so choose the strongest model your computer can run that supports tool use. Smaller models answer simple lookups well but miss more in a full review. If your company runs its own model on a server, contact us and we will help you connect it.
Claude desktop app or claude.ai
Their connectors can't send the X-Esharp-Access header today. Use Claude Code instead; it works with the same Claude account.
Step 3: Verify
Ask your assistant: "Which tools do you have from probe and trace?" It should list tools from both. If it doesn't, restart the assistant and check the addresses and the header. If Probe lists only a handful of tools, the key isn't reaching it.
Step 4: Decide how your assistant asks for approval
Most assistants ask before every tool call, and a review question can take a dozen or more calls to Probe. Whether to approve calls one by one is your decision. Our advice: let the assistant call probe and trace without asking, since those calls only go to your evaluation server, but keep approvals on for editing files and running commands on your computer. Avoid settings that turn off all approvals unless you work in an isolated environment, such as a container or virtual machine. Where to change this depends on the assistant; look for tool or MCP approval settings, and ask us if you can't find them.
How long your key lasts
Your welcome page shows the date your key stops working. When it does, press Create a new key on the welcome page and replace the old value in your configuration, for both servers. With Codex, update the environment variable and restart Codex.
Using an assistant that isn't listed?
If it supports remote MCP servers over HTTP with a custom header, the same addresses and header work. If it doesn't, or you are unsure, contact Daniel Hansson, E-Sharp · daniel.hansson@esharp.se.