Shared team memory for OpenAI Codex

Why this matters in OpenAI Codex

Coding agents forget. The bug your teammate's agent already chased down last Tuesday, the deploy gotcha, the architectural decision from last quarter — none of it is visible to a fresh OpenAI Codex session. threadctx gives every session two tools: memory_query (check the team's memory before starting) and memory_write (save what was learned). What one teammate's agent learns, everyone's agents remember — whether they run OpenAI Codex, Claude Code, Cursor, or anything else that speaks MCP.

Connect OpenAI Codex

Two options, same memory either way:

Local process — register the npm package as an MCP server in OpenAI Codex's MCP configuration (see the Codex MCP docs). Most tools use the standard server block:

{
  "mcpServers": {
    "threadctx": { "command": "npx", "args": ["-y", "threadctx-mcp"] }
  }
}

Remote endpoint — if OpenAI Codex supports HTTP MCP servers with custom headers, skip the local install entirely: https://threadctx.dev/mcp with Authorization: Bearer tctx_…. Details on the connect page.

How the habit reaches OpenAI Codex

MCP tools are pull-based — an agent only calls them if something tells it to. On first start in a repo, threadctx writes a short, clearly-marked instruction into the project's agent rules. threadctx writes AGENTS.md — the instruction standard Codex itself popularized — so Codex picks up the team-memory habit natively.

Set up once, whole team covered

npx threadctx-mcp init          # first teammate, commits the config
npx threadctx-mcp join          # everyone after — one command

init writes committable, secret-free config files; teammates who open the repo in a project-config-aware client are prompted to enable threadctx automatically, and everyone else joins with one command. Local mode is free and makes no network calls; team (cloud) mode shares memory across the team — see pricing and the security page.