Continuo

Continue a coding-agent session in a different agent. Continuo is a macOS menu-bar app that converts a recent Claude Code, Codex, or OpenCode session into another agent's native format and opens it already resumed in your chosen session destination — in any direction, on demand.

Switched tools mid-task? Hitting a model's limits? Want a fresh agent with the same context? Open the picker, pick a session, click the agent you want to continue in.

Install

brew install --cask yoavf/tap/continuo

Or download the latest Continuo.dmg from Releases, drag it to Applications, and launch it — it's signed and notarized.

Recent sessions from ~/.claude and ~/.codex show up automatically. Click the menu-bar icon, or press ⌥⌘S from anywhere, to see them.

Using it

• Pick a session from the menu-bar picker — the newest across all three tools, searchable.
• Choose where to continue it. Click a session to open the continue panel and pick one of the other two agents.
• Choose how much to carry over — the full transcript when it fits the target model's context, or a handoff brief (a structured summary plus your most recent exchanges) when it wouldn't.
• It opens resumed. Continuo writes the target's native session file and opens the resumed session in your chosen destination.

Settings lets you choose where Codex and Claude sessions open, select a terminal app for CLI sessions, pick target models per direction, and toggle the hotkey. Session folders are detected automatically; custom locations and installation controls stay tucked under Storage & installation.

Continuo checks GitHub Releases for cryptographically signed updates using Sparkle. You can check manually from Settings; Sparkle offers background checks using its standard macOS opt-in prompt.

Supported session destinations

A session destination can be a desktop agent, conventional terminal, or workspace environment. Continuo currently supports:

• ChatGPT Desktop (Codex) — open the converted session directly as a local Codex task.
• Claude Desktop (Code) — import the converted Claude Code session and open it in Claude Desktop.
• Terminal, iTerm2, and Ghostty — open the resumed agent directly in a new terminal session.
• CMUX — create a named workspace and run the resumed agent inside it. One-time password access must be enabled under CMUX → Settings → Automation.

How it works

Native session files are read-only inputs. Continuo only ever writes its own mirror files, tracked in bridge-state.json; it refuses to overwrite anything it doesn't own, and never rewrites a mirror you've since continued natively (it makes a fresh one instead). Nothing runs in the background — sessions are converted only when you ask.

Converting a transcript means translating each tool call into the target agent's vocabulary (Claude's Bash ⇄ Codex's exec_command, and so on) so transplanted history reads naturally. Model matching is conservative: the source model is preserved in metadata, and the converted session gets a target-provider model from built-in family matches or your Settings defaults. Hidden provider reasoning is never carried across * — only the visible conversation, tool summaries, and safe metadata.

When a session's title is just a raw prompt, and when it builds a handoff brief, Continuo uses your Mac's on-device model (Apple Intelligence, where available) to write a clean title or a custom compaction of the conversation. Nothing about your sessions is sent anywhere — the on-device model runs locally, and everything else is plain file conversion.

* The latest OpenAI and Anthropic reasoning models return their chain-of-thought as encrypted, provider-owned tokens that never appear in the readable transcript and can't be replayed into another provider. That reasoning is necessarily left behind when you continue a session elsewhere.

What carries over

• Your prompts and the agent's replies, in full
• Tool calls — name, input, and output — remapped to the target agent's equivalent operation
• Timestamps and session metadata: model, working directory, title

What stays behind

• Provider reasoning and thinking traces (encrypted and provider-owned — see above)
• Token counts and billing/usage data
• System prompts and other injected context (the target agent supplies its own)
• Raw file-history snapshots

OpenCode sessions are read from its opencode.db and written through the official opencode import command; models cross in with their provider prefix (anthropic/claude-…, openai/gpt-…) and cross out by stripping it.

Transfer budgets come from each model's real context limit, sourced from models.dev: a snapshot is bundled at build time and refreshed weekly at runtime, falling back to the bundle when offline.

Similar tools

• session-porter — a CLI for porting agent sessions between tools, if you'd rather work from the command line.

Build

swift test
./Scripts/package-app.sh   # → dist/Continuo.app

Signed, notarized release DMGs are built in CI — see docs/RELEASING.md.

License

MIT — see LICENSE. Bundled model metadata is from models.dev, also MIT.