Fix the bad turn. Keep the work that followed.

Agent work became serious before its workspace did.

A coding agent can spend an afternoon inside one thread, another can take a second piece of the problem, and a third can be running on a remote workstation. The work is real, but the operating surface is still scattered across terminal tabs, provider histories, account limits, and project directories.

The expensive failure is not that an agent sometimes gets something wrong. It is that correcting one bad turn often means throwing away everything useful that happened after it, or starting another chat and explaining the project again from memory.

Crowsnest gives that work one lookout: a place to see the history, understand where every run came from, repair supported mistakes, and decide where the work goes next.

Repair one Claude turn without discarding the later work.

Claude Code stores each conversation as linked JSONL records. Crowsnest can rewrite one supported user or assistant text block in place while preserving its UUID, its tool pairings, and every untouched line byte for byte. A timestamped backup is written first.

The result is deliberately narrow: the corrected conversation keeps the work that followed, and claude --resume can continue with the original surrounding context. Rewind and branch remain separate operations for the moments when discarding or comparing later turns is what you actually want.

Codex has a different boundary. Completed Codex history remains immutable; prompt edits and answer retries use the native app-server fork operations instead of rewriting a rollout behind the protocol's back.

See exactly what safe transcript repair preserves →

Claude Code and Codex stay native, but no longer stay scattered.

Crowsnest reads Claude Code history from the files already on disk and reads Codex history through the supported codex app-server protocol. Both appear in one provider-marked gallery with a project tree, stable conversation URLs, title and metadata filtering, and full-message search.

Open a conversation to inspect its messages, tool activity, model evidence, usage, cost estimates, images, and lineage. Start or continue native work without turning either provider into a generic imitation of the other.

Continue across providers, accounts, and machines without pretending they are one thread.

A Claude Code conversation cannot become a Codex thread, and a session stored under one provider account does not magically exist under another. Crowsnest makes those seams visible.

A cross-provider handoff or account continuation creates a fresh native session with an attributed, bounded briefing and a lineage link back to the source. The original history is left unchanged. When several provider accounts are configured, you can compare readable allowance evidence and choose where a new conversation begins; optional routing never treats a missing reading as zero.

Remote Workspaces applies the same principle to location. Connect a machine you already control through one hardened OpenSSH stream. Crowsnest sends and verifies its own helper over that connection, so there is no Crowsnest package to install on the remote machine, no hosted synchronization service, and no remote port to expose.

Parallel work is useful. Reviewed, integrated work is the outcome.

Crowsnest can run independent Claude Code and Codex sessions side by side, but four panes are not a delivery process. Agent Crew starts with a read-only lead turning one objective into an editable dependency graph.

Before anything runs, you can review providers, prompts, dependencies, checks, file claims, and worktree paths. During execution, tasks can be steered, stopped, retried, or rerouted. Integration and independent cross-review are explicit stages; a Crew cannot report Done until both complete.

The point is not to hide more agents behind one button. It is to make their plan, boundaries, decisions, and final integration easier to inspect.

Read the reviewed Agent Crew workflow →

Your conversations stay on the machine that owns them.

The Crowsnest interface binds to 127.0.0.1. It does not upload transcripts, prompts, answers, project paths, conversation titles, usage counts, crash reports, or telemetry.

The official app talks to the Crowsnest account service only for sign-in and session refresh. A current signed access document is verified locally and works offline without payment polling during ordinary use. A Remote Workspace is an SSH connection you explicitly open to your own machine; its conversations remain there.

Claude Code and Codex still contact their own providers when you run them. Crowsnest coordinates those installed runtimes; it does not replace them or proxy their model traffic.

What Crowsnest does not claim to be.

• Not hosted agent execution. Claude Code and Codex run under your own provider plans on a machine you control.
• Not pooled quota. Every provider account keeps its own native configuration, history, and allowance windows.
• Not a synthetic merged thread. Provider and account changes create explicit forks with attributed briefings and visible lineage.
• Not equally editable history. Claude is the deep editable format; completed Codex items remain immutable and use native fork or retry operations.
• Not a team cloud. Local operators are attribution labels, and operations bundles are content-free summaries rather than shared live state.
• Not every platform yet. The official 0.1.8 app is for Apple silicon Macs running macOS 13 or newer.

Start with the complete one-agent workspace.

The first sign-in starts one fixed seven-day Standard trial. Standard covers browsing, search, Claude transcript repair, one live Claude Code or Codex conversation, and the current usage window. Pro adds the multi-agent operating layer: handoff, parallel agents, Crew, UI Studio, Remote Workspaces, multiple provider accounts, full Insights, and local operational controls.

Standard is $19.99 one-time and Pro is $69.99 one-time. Premium is a $99 one-time launch offer that includes every paid module and lifetime updates. Model usage is not included; Claude Code and Codex remain on your existing provider plans.

Written against the app that actually ships.

Capability and privacy claims were reviewed against the Crowsnest 0.1.8 product source and its signed macOS release. The public artifact, checksum, and release notes are available from the Crowsnest 0.1.8 release.