Your coding agents. One live view.

The session transcripts Claude Code, Codex, and OpenCode already write to disk, plus the Git and process metadata needed to tell a running session from a finished one. There is nothing to connect and no account to create — it reads what is already on your machine, plus one Linux host over read-only SSH if you choose to add one.

Claude Code, Codex, and OpenCode — both the desktop apps and the CLIs, with subagents shown under the session that spawned them. coSlash itself runs on macOS, Apple Silicon or Intel; it does not run on Linux or Windows yet. A Linux machine can still be watched from your Mac — see the next answer.

Yes — one Linux host, and only once you add it. In Settings → Machines you give coSlash an alias that is already in your OpenSSH config; it uses the system ssh client and never edits that config. It checks SSH and SFTP, then asks before installing a small collector in the SSH user’s own home directory — no root, no network access, and a fixed allowlist of agent transcript paths. Read-only SFTP is the fallback if the collector cannot run.

Remote sessions land in the same list as local ones, with their host on the card: cards, facts, costs, tokens, file-edit summaries, Resume, and handoff all work while the host is connected. Diffs, synthesis, preview, and command history stay local-only. When the remote process state cannot be confirmed, coSlash labels the session Liveness unknown instead of guessing. Raw remote transcript bytes are never written to your disk — only normalized facts, with no remote paths in them.

No. Transcripts are opened read-only and never written to, and coSlash makes no commits, branches, or edits in your repository. It writes only to its own directory, ~/.coslash, where cached debriefs and pending handoffs live. Resume and handoff launch your own CLI in your own terminal — every edit after that is the agent’s, exactly as if you had started it yourself.

Not unless you turn it on. Debriefs are derived from the transcript with no model involved. Optional AI synthesis only sharpens them, and stays off until you enable and save it. When on, it sends a bounded set of derived facts — about 12 KB — through the local Claude Code or Codex CLI you pick. That spends a little of the agent’s own usage: a few cents per session at most, and only substantial sessions qualify.

No. coSlash binds to 127.0.0.1:8787 and nothing else, and every API request carries an access token minted when the server starts. It is not reachable from your network, and a link from an earlier run stops working once the server restarts.

They are estimates at list API prices, not invoices: token counts from the transcript against a dated per-model price table. A model with no verified price is flagged and left out of the total rather than guessed at, so the number is never quietly wrong. OpenCode sessions report the cost they recorded themselves instead of an estimate.

Yes. Resume reopens the original session in its own CLI with full context. Start fresh writes a Markdown handoff — objective, current state, key decisions, timeline, files, commits, next steps — and opens a new session with it loaded as reference, so the next agent waits for your instruction instead of acting on stale notes. The handoff is plain text, so it crosses vendors: a Claude Code session can be picked up in Codex.