Launched this week

Compagnion
Claude Code & Codex agents side by side, steered from iPhone
5 followers
Claude Code & Codex agents side by side, steered from iPhone
5 followers
Run Claude Code and Codex agents on your Mac, each in its own git worktree. Start from a PR, branch or issue, review changes and open PRs. Let one chat hand off work to new workspaces, each linking back. Give agents your project’s databases, with an access level each. Steer them from your iPhone, on TestFlight.




Free
Launch Team

CloudflareDeploy serverless functions globally in seconds
Promoted
Running Claude Code and Codex side by side in separate worktrees is exactly the setup I've been piecing together by hand. Two things I'm curious about: when one chat hands off to a new workspace, does the context carry over, or does the new agent start fresh? And how do the per-agent database access levels work in practice — read-only by default? Steering from the iPhone is the part that would actually change my day, since most of my check-ins on agents happen away from my desk.
@jcbuilds
thanks! Good questions, so here are the details.
Handoff: the new agent starts fresh. It gets its own worktree and branch, and its first message is a prompt that the handing-off agent writes for it. So the context it gets is whatever that agent puts in the brief. In practice that works well, because the agent handing off knows what matters. Unless told otherwise, the new session keeps the same agent, model, effort and permission mode, but the agent can choose differently. For example, a Claude chat can hand a piece to Codex. The new chat shows "Sent from" with a link back. The original chat can follow up: it sees each workspace's status, PR and latest reply, and can send more instructions. Every agent can also search and read earlier chats, so if the brief misses something, it can look it up in the original conversation instead of you explaining it again.
Database access: the level is set per database connection in a project, not per agent. Every agent in that project, Claude or Codex, goes through the same sql tool. Agents only see the connection's name, never the connection string. Every query is parsed and classified first. There are four levels: always ask, read-only, ask on writes (the default) and read-write. With the default, reads run without asking. Writes, procedure calls and anything the parser can't make sense of ask you first, showing the SQL and a reason like "DELETE without WHERE". Read-only refuses writes outright. Anything that reaches outside the database (xp_cmdshell, COPY … PROGRAM, dblink, GRANT) always asks, even on read-write. These approvals are separate from the agent's permission mode, so "Allow all" doesn't approve them. It supports Postgres and SQL Server for now.
iPhone: that's how we use it too. You follow the transcript live, answer approvals (the SQL ones too) and Claude's questions, send or queue messages, stop a turn, switch the permission mode and start a new workspace from a prompt. "Waiting for you" at the top shows what needs you. The phone connects directly to your Mac over TLS with a pinned certificate, and you pair it with a QR code. On the same network it just works. Away from home you need a route to your Mac, such as Tailscale on both devices. Not there yet: push notifications while the app is in the background, and a relay so you don't need Tailscale. Both are on the list.