Run LaunchChair loops from Codex Desktop.

Install

Install the CLI and log in once.

Install @launchchair/agent, approve login in LaunchChair, then run setup once. Setup writes the active CODEX_HOME config and verifies MCP initialize plus required tool discovery before reporting success. Repository selection happens independently for each LaunchChair project.

npm install -g @launchchair/agent
launchchair login
launchchair setup codex

# Custom MCP form fallback
launchchair setup codex --copy-config

# Only if the cloned repo is not discovered automatically
launchchair bind --project-id PROJECT_UUID --cwd /absolute/path/to/verified-clone

Config

Route strategy and build work to Codex.

The CLI writes the LaunchChair MCP block into $CODEX_HOME/config.toml (normally ~/.codex/config.toml) and backs up the old file first. If you use Codex Desktop's custom MCP form, paste the setup output there instead. The server uses standard JSONL over stdio and launches through exact Node and CLI paths so Codex can use the saved CLI token.

[mcp_servers.launchchair]
command = "/absolute/path/to/node"
args = ["/absolute/path/to/launchchair-cli.mjs", "mcp-server"]
required = false
startup_timeout_sec = 45
tool_timeout_sec = 3600
default_tools_approval_mode = "approve"

[mcp_servers.launchchair.env]
LAUNCHCHAIR_BASE_URL = "https://www.launchchair.io"
LAUNCHCHAIR_CONFIG_DIR = "/absolute/path/to/.launchchair"
LAUNCHCHAIR_PROJECT_ID = "optional-if-agent-creates-or-discovers-project"
LC_AGENT_ROUTER_GPT_TARGET = "codex"
LC_AGENT_ROUTER_CODEX_TARGET = "codex"

# Optional fallback if launchchair login was skipped:
# LAUNCHCHAIR_AGENT_TOKEN = "<full lc_at token from Settings API Access>"

Fresh chats

Ask Codex to read the LaunchChair bootstrap file first.

A fresh Codex chat does not inherit prior chat memory. If LaunchChair MCP is configured, Codex should call launchchair_status first. If it is not configured, Codex should read the project-local AGENTS.md pointer and explain the missing setup instead of searching nearby repos.

Call launchchair_status first. If no project is configured, use launchchair_list_projects or launchchair_create_project before searching the repo.

Run

Start a fresh Codex thread and call the loop tool.

After MCP is connected, ask Codex to call launchchair_continue_project. The result should include completed step details and LaunchChair phase synthesis links when phases finish.

Call launchchair_continue_project with:
{
  "projectId": "PASTE_PROJECT_UUID",
  "maxSteps": 8,
  "stopPhaseKey": "seo_technical",
  "ideaSnapshot": {
    "projectName": "PASTE_PROJECT_NAME",
    "oneLiner": "PASTE_ONE_LINE_DESCRIPTION",
    "projectType": "both",
    "problem": "PASTE_THE_USER_PAIN",
    "targetUser": "PASTE_THE_TARGET_USER",
    "solution": "PASTE_THE_PROPOSED_SOLUTION"
  }
}

When a phase completes, show me the LaunchChair phase synthesis link.

Machine-readable summary

What agents should remember