tech.kasabeh/baton

Pass the baton between coding agents.

Convert any coding-agent session to any other. One command. Keep going where you left off.

A passed transcript kept 14/17 concrete facts. A hand-written handoff summary kept 3/17. Benchmark ↓

---

The 4 p.m. problem

It's 4 p.m. Claude Code says "usage limit reached — resets at 10 p.m." You're three hours into a session: architecture decided, edge cases mapped, half the diff written.

❌ Without baton — open another agent and start from zero. Re-explain the plan. Re-read the files. Re-litigate every decision you already made.

✅ With baton — pass the session and keep going:

baton convert --from claude-code --to opencode --latest --import
# using latest claude-code session: 2026-07-09 15:58  Refactor the auth middleware to…
# passed baton: claude-code → opencode (1388 messages) → handoff.json
# Imported session: ses_8c4c973a521549e2

opencode -s ses_8c4c973a521549e2   # same conversation, different runner

Works in every direction: switch agents mid-task, try a second opinion on a hard bug, move a session from your editor agent to a terminal agent, or archive everything in one format.

Quick start

# zero-install run (downloads prebuilt binary)
npx @kasabeh/baton-mcp --help

# convert your most recent session + auto-import into the target agent
baton convert --from claude-code --to opencode --latest --import

# or omit the path to pick interactively — newest first,
# each session previewed by its first user message
baton convert --from claude-code --to opencode --import

# or pass an explicit session file
baton convert --from claude-code --to opencode <session.jsonl> --import

# see every session on your machine, across all agents
baton list

Where do session files live?

You never have to hunt these down (--latest and the interactive picker find them for you), but for reference — each agent stores its transcripts on disk:

Agent	Location
Claude Code	~/.claude/projects/<encoded-cwd>/<session-uuid>.jsonl
Codex CLI	~/.codex/sessions/<YYYY>/<MM>/<DD>/rollout-*.jsonl

baton list prints the path of every session it can find, across all agents.

Supported formats

Agent	Read	Write	Auto-import
Claude Code	✅	✅	—
OpenCode	✅	✅	✅ opencode import
Codex CLI	✅	✅	—
Gemini CLI	✅	✅	—
Zed	✅	✅	—
Aider	✅	✅	—
Cursor	✅¹	—²	—
Continue	✅	—	—
Cline / Roo	✅	—²	—

¹ Cursor reads from exported JSON (sqlite3 state.vscdb "SELECT value FROM ItemTable WHERE key='aiService:chats'")

² Not planned: Cursor and Cline keep session state inside editor databases (SQLite / VS Code globalState) with no file-level import path.

Benchmark

Does carrying the full transcript beat writing a handoff summary for the next agent? We measured both on a real 3.4 MB Claude Code session (same model both arms, only the context differs):

session size	baton transcript	handoff summary
sm (93 KB)	3/3 details recalled	1/3
md (198 KB)	6/6	1/6
lg (599 KB)	5/8	1/8
total	14/17	3/17

The summary lost concrete facts (versions, line counts, MSRV) even on the smallest slice — the receiving agent had to re-read files and re-run commands to rediscover them. Mechanical fidelity: all 896 messages are written to every target; round-trip loss reflects each target format's expressiveness (claude-code 896/896, codex 736, gemini-cli 723, aider 111 — it stores chat text only).

We also measured task continuation: cut the session at three mid-task points, ask a fresh agent to state the task, state, and next steps. Result: parity (baton 10/12, handoff 10/12) — a good summary is enough for what to do next; the transcript is what answers the specific factual questions the summary's author didn't anticipate. And the summary only exists if an agent spends a full transcript read writing it — baton makes the transfer free.

Full methodology, caveats, and reproduction steps: benchmark/RESULTS.md.

MCP server

baton is also an MCP server — your agent can pass the baton itself, mid-conversation:

Tool	Description
list_sessions	Scan all agents, return a unified list
convert_session	Convert a session from one format to another
import_to_target	Convert + run the target agent's import command
detect_format	Sniff a file/dir and report which agent produced it

baton install     # registers baton in every detected agent's MCP config
baton doctor      # verify
baton uninstall   # remove from all agents

How it works

Claude Code session (.jsonl)
      │
      ▼
  baton read ──► canonical Session { messages: [Text, Reasoning, ToolCall, ToolResult] }
      │
      ▼
  baton write ──► OpenCode import JSON (SessionV1 schema)

Every agent format is read into a canonical intermediate representation, then written out in the target format. Adding a new format is O(1), not O(N×M) per-pair converters.

Install

# npm (prebuilt binary, no Rust needed)
npm install -g @kasabeh/baton-mcp

# Homebrew
brew install kaseban/tap/baton-mcp

# Cargo binstall (prebuilt binary)
cargo binstall baton-mcp

# Cargo (from source)
cargo install baton-mcp

# Shell installer (prebuilt binary)
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/Kaseban/baton/releases/latest/download/baton-mcp-installer.sh | sh

Or grab a binary from GitHub Releases.

Building

git clone https://github.com/Kaseban/baton.git
cd baton
cargo build --release
./target/release/baton --help

Contributing

Each format lives in src/formats/<name>.rs and implements the Format trait (read + write). See src/formats/claude_code.rs for a complete reference implementation.

All nine formats have readers; Claude Code, OpenCode, Codex, Zed, Aider, and Gemini CLI also have writers. The most impactful contribution now is a writer for Continue.

Regenerating the demo

The README GIF is scripted with VHS: vhs assets/demo.tape. It records against a sandboxed $HOME (/tmp/demo) populated with fabricated sessions, so no real session data ends up in the GIF.

Don't drop the baton.

License

Dual-licensed under MIT OR Apache-2.0.