Handoff
Let your AI agents work as a team on the same project.
Say "Codex, take the tests; Claude, do the API" in either app. The work, notes and reviews pass between them
through a shared task board, so you stop being the copy-paste courier. Or hand out a whole list at once, to any
model you use:
handoff dispatch "codex review my changes, grok make pictures of my app, claude finish the next step"
Handoff works in Claude Code, Codex, Claude Desktop, Gemini CLI, OpenCode and Grok Build. The board is a file
in your project, and Handoff never listens on the network.
ixelai.com/handoff · Docs ·
Privacy
What it is, and what it isn't
- No account, no telemetry and no server of ours. Handoff never listens on the network. An agent run sends its
task to that agent's company, a run on a local model talks to that model's server, and
handoff update and
handoff rankings update reach GitHub.
- It never pushes, and it never merges on its own. You merge a finished task yourself, with
handoff merge or
the Merge button in the Ixel app, into the branch you have out.
- An agent runs a task only after you approve it, and an approval is good for one run. Handoff doesn't pick up
work on its own.
- It isn't a model or a hosted service, and it calls no model itself. It runs the agents, subscriptions, API
keys and local models you already have.
- The sandbox that lets an agent run your tests and builds is bubblewrap on Linux and the built-in
sandbox-exec on macOS. On a Mac, commands the agent runs can still read the environment of your other
programs (macOS allows it and its sandbox can't stop it), so keep tokens out of your shell's environment
while a worker runs; SECURITY.md has the details. On Windows, Claude Code, OpenCode and local models only
edit files for now. Kimi,
Cline and OpenClaw run only once you enable them, and then with their own permissions: commands as you,
with the network.
Install
Windows, in a normal PowerShell window (not "Run as administrator"):
irm https://ixelai.com/handoff/install.ps1 | iex
macOS or Linux:
curl -fsSL https://ixelai.com/handoff/install.sh | sh
You need Python 3.10 or newer and Git. To get Ixel MAT and the Ixel app too,
install all of Ixel instead. To update or uninstall, see the
install guide.
Then open a new terminal, go to your project's folder, connect your apps and restart them:
handoff setup --write
Use it
- In any connected app, ask the agent to hand work over. In the other app, say "check your handoffs".
handoff board shows who's doing what. handoff 3 shows task T-3, and handoff done 3 closes it.handoff dispatch "..." splits one request across your agents and runs the parts side by side.handoff approve 3 opencode (or claude, codex, ollama, kimi…) lets one agent work on T-3 in its own
branch. Every coding agent and local model you have can take tasks: handoff agents lists them.
Every command, tool and rule is in the guide.
Models, chains and the worker
Once you approve a task, the worker runs the agent on it in its own git
worktree and branch. It never pushes. You choose which models do the work:
handoff approve T-7 claude --model claude-haiku-5-5
handoff approve T-7 claude --chain claude-opus-5-5,claude-sonnet-5-5,ollama:qwen3:4b-instruct
- A model, or a chain. A chain lists models biggest first. The first plans up to 6 steps and the smaller
ones do them, each step committed on its own on the task's branch. A step can go to another agent as
agent:model, such as ollama:qwen3:4b-instruct. Work comes back up the chain: the biggest step model checks
the smaller models' work and fixes it, and anything too big goes up to the planner.
Details.
- Effort per model, in each agent's own words:
claude-opus-5-5:high in a chain, or --effort (claude takes
low to max, codex minimal to ultra).
- What it used. A run reports the tokens each model used and an estimated cost at API prices.
handoff show T-7 and the handback show it. A local model reports its tokens at no cost.
- Merge.
handoff merge T-7 merges a finished task's branch into the branch you have out, as one merge
commit. Nothing is pushed. The task is marked done and its worktree removed. It refuses, changing nothing,
when the merge could touch files git doesn't track or the board.
- Usage limits. When a model hits its usage limit, the next best model takes over, and the handback says so:
"claude-opus-5-5 hit its usage limit, so claude-sonnet-5-5 took over." Approve shows you who would take over
first.
--no-fallback turns it off. Details.
- Rankings.
handoff rankings shows which models are best at what, in tiers. Top plans and reviews
(claude-fable-5-1, claude-opus-5-5, gpt-6-astra). Middle builds and checks (claude-sonnet-5-5, gpt-6-sol).
Small takes short, exact directions (claude-haiku-5-5, qwen3:4b-instruct).
Recommended setups has the rest.
- Other agents. OpenCode, Ollama, LM Studio, Hermes, Kimi, Cline and OpenClaw can take tasks too. Hermes,
Kimi, Cline and OpenClaw run only once you enable them.
What each can reach and cost.
- The Board in the Ixel app. The Board is also in the Ixel app (, installed with
); from
shows the same pages in a browser. There you can approve runs with a model picker, see what each run used and
cost, and merge with the Merge button.
On Windows the worker is edit-only for now. Claude Code, OpenCode and local models change files but run
no commands, tests or builds, and Codex doesn't run. On Linux (with bubblewrap) and macOS (its built-in
sandbox-exec, nothing to install), the agent runs in the sandbox that runs your tests and builds.
If a run fails
handoff doctor shows whether each worker is ready, and whether it found that agent's login.handoff show T-N shows the task's history, including why a run failed.- A run that finds no login doesn't start, and the task stays approved: sign in, then
handoff run T-N.
- The guide's Troubleshooting section has more.
Privacy
Handoff has no account, no telemetry and no network code of its own. The apps you connect read the board, and
an agent run sends its task to that agent's company. Some companies train on what you send:
which ones, and how to keep your work out.
More
License
MIT
