Wiring your editor

WhyGraph reaches your editor over HTTP MCP. Each project has two endpoints on the portal: http://127.0.0.1:<port>/mcp/<slug>, WhyGraph's own tools, and http://127.0.0.1:<port>/mcp/<slug>/code, CodeGraph's graph tools served by the portal. Any agent that speaks MCP over HTTP can use them. You don't write that config yourself: when you set up a project, you pick the agents and the portal writes each one's entry, after showing you a preview.

Supported agents

Four agents are supported. All of them are project-scoped - the config file is written or merged inside the repo, so you can commit it and every teammate's editor picks it up.

Agent	Editor	Config file	Assets land in
claude	Claude Code	.mcp.json (repo root)	.claude/
cursor	Cursor	.cursor/mcp.json	.cursor/
vscode (alias copilot)	VS Code / GitHub Copilot	.vscode/mcp.json	.github/
codex	OpenAI Codex	.codex/config.toml	repo root + .codex/agents/

You can select several agents for one project, and change the set later in the project's Settings.

What to know

The details live on Connecting agents:

• the exact entry each agent gets, and why it is safe to commit on the default port;
• what to do after a non-default port (export WHYGRAPH_PORT for Claude Code, edit the literal port for Cursor and Codex, enter it in VS Code's prompt);
• Claude Code's approval prompt for project-scoped servers, Codex's trusted-project rule, and VS Code listing the server twice when .mcp.json is also written;
• migrating a 1.x entry that still runs whygraph-mcp, which no longer exists.

Bundled assets

Every agent gets an asset tree, not just Claude Code - subagents, commands, and skills that teach your editor how to use WhyGraph's tools. Re-initializing leaves your existing files alone; Update agent files in the project's Settings overwrites them.

The planning commands take flags: /whygraph-plan accepts --shallow, --deep, --split, --no-split, --sources <a,b>, --no-questions, --review, --no-review and --out <path>; /whygraph-implement accepts --from-step N, --step N, --commit-each-step, --fix-attempts N, --model, --no-review, --no-eval, --no-capture and --eval-fix-attempts N. See Planning a change.

Cursor's slash commands (/rationale, /whygraph-plan, /whygraph-implement, /whygraph-ui-check, /whygraph-note) are installed as skills under .cursor/skills/. Cursor also loads .claude/skills/, so with both Claude Code and Cursor selected each command shows up twice in Cursor's slash menu. A project initialized by an earlier build has old .cursor/commands/ files: Update agent files removes them (it asks first if you edited one). Copilot also reads AGENTS.md, so with both VS Code and Codex selected the two always-on blocks overlap.

Each agent's install also append-merges a CodeGraph usage-guidance block, pointing at the portal's codegraph_* tools, into that agent's always-on instructions - CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md, and an always-apply rule for Cursor. Your own content is preserved; the block is added below it.

Verify

Check the portal is running (whygraph status), then ask your agent to list its MCP servers - whygraph and codegraph should both be there. Next, see how an agent actually calls the tools.