Coming from Codex or Claude Code

You don't start over when you move to agentty. Your past Codex and Claude Code conversations show up in the thread list, and you can open one and keep talking. Your instructions, skills and slash commands are read from where those tools keep them. MCP servers come over with one command.

Nothing here changes the other tool's files. You can keep using both.

Continue a conversation

Open the thread list with Ctrl+J. Codex and Claude Code sessions sit next to your agentty threads, newest first. They're marked with ↗ and the tool's name:

▎ ● fix the windows build                                  3 msgs · now
  ↗ what did you do with the agentty? It's not opening.    Codex · 75 msgs · 3mo
  ↗ Skip to content 1ay1 agentty Repository navigation…    Codex · 212 msgs · 3mo

Press Tab to cycle through all → this folder → Codex & Claude, and S-Tab to go back. Tab skips any list that would be empty. The panel title shows which list you're on, and the hint row shows where Tab goes next.

With Thread preview on (Appearance), the card above the list shows the session's turns. Replies are labelled with the tool that wrote them, and the hint reads Enter continue here.

Press Enter and the session opens as an agentty thread. Type your next message and send it. The model gets the whole conversation, so it knows what you were doing, even if it's a different model from a different company. A line marks where agentty took over:

┈┈┈┈┈┈┈┈┈┈┈┈┈┈  → continued in agentty · from Codex  ┈┈┈┈┈┈┈┈┈┈┈┈┈┈

What comes across

What doesn't come across, and why

• Encrypted or signed reasoning. Codex stores its reasoning encrypted, and Claude Code signs its thinking. Both only work for the account and model that made them, so they're dropped. The readable part stays.
• Sessions that weren't yours. Runs started by other programs (codex exec, SDK and editor integrations) and Claude Code's subagent side-chats aren't listed.
• Rewound branches. If you rewound a Claude Code conversation, you get the branch you ended on, the same thing Claude Code shows.

How tool calls look

Each call is mapped to the agentty tool that does the same job, so you get the same cards you'd get from an agentty turn:

For Codex patches, the diff shown is the one Codex actually applied, not just the patch it asked for. Anything else keeps its own name and arguments and shows as a generic card. Codex runs its tools through small scripts. A script that makes one tool call shows as that tool; a script with several shows as Codex Script.

When you keep using the other tool

Say you open a Codex session in agentty, ask a couple of things, then go back to Codex and carry on there. Next time you open it in agentty, it re-reads the session: Codex's new turns come in, and your agentty turns stay after them.

Opening the same session again always goes back to the same thread. You never get duplicates.

Where it's saved

Opening a session saves it as a normal thread under ~/.agentty/threads/. From then on it works like any other thread: rename it with Ctrl+R in the thread list, fork it, compact it, delete it with Ctrl+D in the thread list. Deleting it in agentty doesn't touch the original session.

Find sessions from the command line

agentty sessions

lists every session agentty can see, newest first:

  0  Codex        212 msgs   3mo ago  Skip to content 1ay1 agentty Repository navigation…
      /home/you/projects/agentty
  1  Codex         75 msgs   3mo ago  what did you do with the agentty? It's not opening.
      /home/you/projects/agentty

1 other file(s) skipped (subagent or SDK runs, empty)

agentty sessions --dump N prints how session N reads once it's an agentty thread: each turn, each tool call with its arguments and output, and where compactions are. It's handy when a session doesn't look the way you expect.

Where agentty looks

The list is read in the background each time you open Ctrl+J. A file that hasn't changed since the last look isn't read again, so a long history costs nothing after the first scan.

Don't want them in the list? Turn off Appearance → Codex & Claude sessions. Threads you already opened stay, since they're agentty threads now.

Your instructions come with you

agentty reads guidance files from the same places those tools do. There's nothing to copy.

Details and the order they're applied in: Configuration → CLAUDE.md and AGENTS.md.

Skills, subagents and slash commands

These are plain markdown files, and agentty reads them where Claude Code keeps them: .claude/skills/, .claude/agents/ and .claude/commands/, in the project and in your home folder. The shared .agents/ folders work too. They behave the same, $ARGUMENTS included. See Skills and Slash Commands.

MCP servers

MCP servers start programs, so agentty doesn't run another tool's config behind your back. It copies the servers over once, after showing you what it found:

agentty mcp import                 # what each tool offers
agentty mcp import --from codex    # copy Codex's servers
agentty mcp import --from claude   # copy Claude Code's servers

--dry-run shows what would change, and --force replaces servers agentty already has with the same name.

From Codex, [mcp_servers.*] in ~/.codex/config.toml (and .codex/config.toml in the project) is translated to agentty's format: command, args, env, working directory, URL, headers, timeouts, disabled tools and the off switch. Two things don't come over, and the import tells you which server they belong to:

• Secrets held in environment variables (bearer_token_env_var, env_http_headers). agentty stores headers as written, so it won't copy a token into a file for you. Add the header by hand, or use agentty mcp-login if the server supports OAuth.
• enabled_tools. agentty filters with an exclude list only.

More in MCP.

Sign-in

Your Claude Pro/Max or ChatGPT plan works in agentty too. agentty login signs in with the same accounts. agentty doesn't read the other tools' saved logins: if two programs share one refresh token, one of them gets signed out. See Authentication.