Connect Codex or Claude Code
Connect your own Codex or Claude Code to an Office’s research workspace and participate directly without a Subject or Embodiment. MCP connects tools to external capabilities. This connection exposes research only.
Prepare your account and key
Install Codex and sign in to your own OpenAI account, or install Claude Code and sign in to your own Anthropic account. The Office must have research enabled and its workspace open, and you need an active research key. The computer running your tool must reach the HTTPS address provided by Connect tool.
- Open Edgewright and select Connect tool.
- Choose Codex or Claude Code under Your tool.
The drawer shows environment-variable commands and tool configuration. If the public address is missing or incorrect, ask the operator to fix it. Displayed settings alone do not establish connection.
Give your tool the key
Set the key in the terminal that starts your tool. Configuration references only the variable name, not the key value.
-
Under Set your key, select Copy command and run it in Bash or Zsh:
read -rs SOULHALL_RESEARCH_KEY && export SOULHALL_RESEARCH_KEY -
Paste the key and press Enter. It is not displayed or written to shell history.
-
Select Show configuration, then Copy configuration.
- Codex: add it to
~/.codex/config.tomlor.codex/config.tomlin a trusted project. - Claude Code: save it as
.mcp.jsonin the project root. Approve the project server when prompted.
- Codex: add it to
-
Save, then start or restart your tool from the same terminal where you set the key.
The tool now has a connection address and a way to read the key. Verify inside the tool next; advancing in the drawer is not enough.
Do not put the key value in configuration files, command-line arguments, or conversations. Closing the terminal loses this environment variable. Set it again before starting next time.
Confirm the connection works
- In Codex, enter
/mcp. In Claude Code, enter/mcpor runclaude mcp listin the terminal. - Check
edgewrightis connected and its tools are loaded, not merely named in the list. - Ask the tool to list workspace datasets. Start with a read-only query before making changes or evaluating.
Loaded tools and a query response establish that the connection works.
| What you see | Next step |
|---|---|
| Connected with tools loaded | Start researching. |
Only the name edgewright | Configuration exists; keep checking connection state. |
| Awaiting approval | Approve the server, then check again. |
| Connection failed or authentication required | Check the environment variable, address, and key state. |
Loaded tools do not mean a backtest completed. Read the operation and evaluation results to establish success. A key’s Last used does not replace that check.
Confirm scope and costs
A key grants read-write access to one Office’s research workspace: upload data, write strategies, run evaluations, and read results. You share it with colleagues and Embodiments; it is not a private copy.
The key does not expose Office conversations, memory, browsers, or reports. It does not turn your tool into an Embodiment or grant Subject knowledge.
Your model account pays your tool’s costs, not Soulhall. Its research does not count toward the Office’s model-turn budget. There is no separate research allowance or research usage bill here.
Research safely
- Protect keys like passwords. If one may have leaked, Revoke it immediately in Edgewright.
- Agree on data and strategies you may change in the shared workspace and respect source usage rights.
- Require the tool to use Edgewright first for formal evaluations. If it cannot, ask it to explain, obtain your agreement to a specific alternative, and label results “Not produced by Edgewright.”
These are research instructions, not a policy Soulhall can force external tools to follow. Agreeing to an alternative does not permit bypassing authorization, safety limits, or engine refusals.
Revoking a key or disabling Office research blocks later requests but does not stop evaluations already submitted.