Codex for LinkedIn outreach

OpenAI Codex is a coding agent, not a LinkedIn client. If you already run it against the repo, the same session can talk to Omentir over MCP, score a batch, and leave drafts. It should not hold a LinkedIn password. It should not invent a campaign while you are midway through a refactor.

This is not the ChatGPT connector. ChatGPT uses Settings, Connectors, and workspace approval. Codex reads ~/.codex/config.toml. Mixing those paths is how people paste a token into a chat UI, or wait for an OAuth screen that never appears.

Why the TOML file matters

The CLI and the IDE extension share the same file. Get the table name wrong (mcp-servers instead of mcp_servers) and Codex ignores the block. If you paste the actual token into bearer_token_env_var, Codex treats that string as a variable name and auth fails. Put OMENTIR_API_KEY there, and export the key before you launch. Then run /mcp in the session. If the server is missing, fix that before you invent a second connect path.

A project-local `.codex/config.toml` only loads for trusted projects. If tools never appear, start with the global file. Setup: Codex integration.

The config Codex expects

Create a revocable key. Export it as OMENTIR_API_KEY. Point url at https://omentir.com/api/agent/v1/mcp. Keep send behind review.

Paste into config.toml

[mcp_servers.omentir]
url = "https://omentir.com/api/agent/v1/mcp"
bearer_token_env_var = "OMENTIR_API_KEY"

A first Codex job

Finish Omentir first. Fetch agents.md. Start with get_context and list_agents. If you want a new finder, ask it to show the config and wait. Replace the brackets. Keep the last two sentences.

Paste into Codex

You are Codex, running against my repo through Omentir.

My product helps [buyer] get [result]. Best-fit accounts are [company type, size, region]. Strong signals: [signal]. Skip [bad fit].

Work only through the Omentir MCP server named in config.toml. Read https://omentir.com/agents.md if you have not. Start with get_context, then list_agents. If no finder matches this ICP, show me a create_agent config and wait for a yes.

Do not put secrets in config.toml. bearer_token_env_var is the name of an environment variable, not the token.

Pull up to 30 people. For each one write fit 1-5, the evidence, any risk, and a two-sentence draft that cites a real trigger. If a note could fit two buyers, rewrite it.

Do not send. Do not enroll. Do not sign into LinkedIn.
Do not create, delete, or reply unless I ask in this session.

Not ChatGPT chat

If you wanted a conversation this afternoon with no repo open, use the ChatGPT integration. If you wanted the editor instead of Codex, use Cursor for LinkedIn outreach. Overnight research is Grok Bot.

When to stay in the browser

You are not in Codex. The product story is still a slogan. You cannot give the list fifteen minutes. Start in Overview. Codex earns the key when that agent is already in the repo.

FAQs

No. ChatGPT chat uses Connectors and workspace approval. Codex uses config.toml and a Bearer env var. Same Omentir tools. Different machine.

Usually ~/.codex/config.toml. A project .codex/config.toml only loads for trusted projects. The table name is mcp_servers, with an underscore.

bearer_token_env_var wants the variable name, not the token. Pasting the key into the file is how it lands in git.

No. Overnight on a shared computer is Grok Bot, a different product.

Use Cursor if that is already the editor. Use Codex if you already run the Codex CLI or extension. Do not run both on day one.

No. Overview already finds people and drafts notes. Add Codex if that agent is already in the repo.