seshport

Port your session between coding agents. Type /seshport inside Claude Code or Codex, open the other agent, and resume the exact same conversation — full context included.

By Harshil-Jani · MIT

Demo

A Codex session's haiku, recited by Claude Code after the port. A Claude Code session's fizzbuzz, recalled by Codex. Real recording, synthetic sessions — nothing staged.

Install

One line — installs the binary and the /seshport slash command for both agents:

curl -fsSL https://raw.githubusercontent.com/Harshil-Jani/seshport/main/install.sh | sh

Or grab just the binary from wherever you live:

npm i -g seshport                        # Node
pip install seshport                     # Python
cargo install seshport                   # Rust
brew install harshil-jani/tap/seshport   # Homebrew

Then copy commands/ into ~/.claude/commands/ and ~/.codex/prompts/ for the /seshport slash command. Prebuilt binaries for mac (arm/x64), linux (arm/x64), and windows are on Releases.

The easy way: /seshport

Never leave your agent. Mid-conversation, just type:

/seshport

• Inside Claude Code → replies with codex resume <id>
• Inside Codex → replies with cd <project> && claude --resume <id>

Open the other agent, paste, and continue the exact same conversation. Verified: a codeword planted in a Claude Code session was recalled by Codex after a /seshport round-trip.

CLI usage

The source tool is always auto-detected; name the target (or omit it when only one target is possible).

seshport <session-id> <tool>   # port any session to any tool (claude | codex | grok)
seshport <session-id>          # auto: found in any tool, only one possible target
seshport <path.jsonl> <tool>   # format detected from file content
seshport claude codex          # newest Claude Code session -> Codex
seshport grok claude           # newest Grok Build session -> Claude Code

Each run prints the output path and the exact resume command:

$ seshport codex
/Users/you/.claude/projects/-your-project/1b2c3d4e-....jsonl
resume with:  cd /your/project && claude --resume 1b2c3d4e-...

Want to try without touching your real sessions? The demo transcripts from the GIF are in demo/:

seshport demo/codex-session.jsonl claude          # -> claude --resume <id>
seshport demo/claude-session.jsonl grok           # -> grok --resume <id>
seshport demo/grok-session/chat_history.jsonl codex   # -> codex resume <id>

Architecture

Every agent is one implementation of the Tool trait against a neutral Transcript — so N tools cost 2·N converters instead of N². The diagram is editable: open docs/architecture.excalidraw at excalidraw.com.

How it works

• User and assistant messages transfer as-is into a neutral Transcript.
• Tool calls/results are flattened to readable text ([tool call: Bash] ...) — provider-specific API state (tool-call ids, encrypted reasoning) can't be replayed cross-provider, but the resumed agent gets the full story as context.
• Thinking/reasoning blocks are dropped (provider-internal).
• Codex output borrows base_instructions from your newest real rollout — ChatGPT-auth Codex rejects sessions without the official instructions.
• Every import starts with an attribution message noting the source session and this tool.

Add your editor — PRs welcome

Every integration is one Tool impl (five methods) plus a demo fixture — see CONTRIBUTING.md for the step-by-step guide, and PR #1 (Grok Build) for a worked example you can copy.

License

MIT © Harshil Jani