mend

The CLI for Mend — a local-first workbench for developers who use coding agents heavily. Adopt a repository into Mend's central store, run your agent (Claude Code, Codex, or any command) in a recorded per-session git worktree, detach, and reattach from any terminal.

npm install -g @sealant/mend

Requires Node 22+ and a running Mend server.

Run bare mend for the interactive dashboard. Highlight a session and press v to review its accumulated change in the terminal: navigate files and hunks, switch between unified and split diffs, reveal whitespace, select line ranges, add line or whole-change comments, inspect linked evidence, and draft a follow-up for the same session. Press y to deliver a pending follow-up and relaunch that session, or o to continue the review in the web app.

Getting started

Once per machine, in this order:

mend login                       # sign in to the server; the token is saved 0600
mend connect codex               # send this machine's codex (or claude, github) credential
mend adopt                       # adopt the repository you are standing in
mend codex "fix the flaky test"  # new session worktree, harness running in it
mend pair                        # hand a phone the same server
mend doctor                      # every fact above, on one screen

mend help prints the same sequence under start, then everything else.

mend pair

mend pair [--url <base url>]

Asks the server for a pairing code and prints it three ways: a QR of the mend://pair deep link, the code grouped as ABCD-EFGH, and the base URL the device should reach — the machine's tailnet address when it has one, otherwise a LAN address (--url overrides the choice). Scan it in the Mend app, or type the URL and the code in by hand. The code is single use and expires in 10 minutes; the device's own token is minted when it claims the code, and can be revoked from the Mend app later. The device names itself when it claims the code.

mend doctor

mend doctor

Reads, and changes nothing: server reachable, token accepted, the Sealant connection, each connected account, adopted projects, the claude / codex / gh CLIs on PATH and whether their credentials exist on this machine, and the tailnet address. One line per fact — ✓ observed, ○ not set up yet, ✗ a blocker — and every line that needs an action ends with the one command that takes it:

✓ server      http://localhost:3105 · mend 0.5.0
✓ signed in   token accepted
✓ sealant     connected · http://127.0.0.1:4000
✓ claude      connected · [email protected]
○ codex       not connected → mend connect codex
✓ projects    2 adopted
○ gh cli      on PATH · no credential here → gh auth login
○ tailnet     not detected

It exits 1 when a ✗ is printed, so a setup script can gate on it. No request waits longer than 3s.

Commands

mend adopt [source] [--name <name>]   adopt a repository into the store (default: cwd)
mend codex|claude|opencode            new session worktree + launch the harness in it
mend run -- <command...>              same, with an arbitrary command
mend attach <session-id-prefix>       reattach this terminal to a running session
mend continue [session-id]            resume a session with its pending review follow-up
mend resume [session-id] [--with h]   rejoin a settled session (state restored; --with switches harness)
mend rejoin [session-id] [--harness h] attach if live, otherwise resume; newest live wins
mend sessions [--all] [--project p] [--json]
mend status                           active sessions (alias of mend sessions)
mend pair [--url <base url>]          pair a phone or a second machine: QR + code + URL
mend doctor                           read-only checklist of this machine's setup
mend skills [--project [p]]           your skill library on the server (or a project's)
mend skills push [--project [p]] [--prune] [--dir <path>]
                                      upload ~/.agents/skills bundles; sessions receive
                                      them at launch (--prune removes what's gone)

Signing in

mend login                 # opens the browser at <server>/authorize; press Authorize there
mend login --url https://mend.example.com
mend logout                # revokes this terminal's device token and forgets it

mend login opens the browser at <server>/authorize?code=… and waits, using the server already configured (--url, MEND_URL, or the config file). Only a fresh machine with nothing set asks for the URL, with Enter accepting http://localhost:3105. Sign in there if needed, check that the code on the page matches the one in the terminal, and press Authorize. The CLI receives a device token of its own, the same revocable kind a paired phone holds, listed under Settings → Devices. No password is ever typed into the terminal. Over SSH the browser does not open by itself; open the printed URL on any signed-in device instead.

The token is stored 0600 in the CLI config below; every command uses it until mend logout. On a dev instance with MEND_STATIC_TOKEN set, MEND_TOKEN=<that value> also works.

Configuration

| Source | What | | ------------------------- | ---------------------------------------------------------------------------------------------------------- | | MEND_URL | The Mend server (default http://localhost:3105) | | MEND_TOKEN | Bearer token for that server (normally written by mend login) | | MEND_DETACH_KEY | Set to none when an outer multiplexer detaches | | ~/.config/mend/cli.json | { "url": ..., "token": ..., "deviceId": ... } — env vars win; a pre-XDG ~/.mend/cli.json keeps working |

Herdr

When Mend attaches Codex, Claude Code, or OpenCode inside a Herdr pane, it reports that harness to Herdr for the lifetime of the attachment. The session therefore appears in Herdr's Agents sidebar whether it was started from the mend dashboard or a one-shot CLI command. Mend supplies Herdr's foreground-process hint rather than claiming lifecycle authority, so Herdr continues to derive working, idle, and blocked from its native agent screen rules.

How it works

Every session runs in its own git worktree over the adopted repository, inside a recorded workspace. The terminal you see is a held WebSocket to that workspace's PTY: Ctrl+] detaches and the session keeps running; mend attach reconnects with full scrollback — from this machine or any other that can reach the server. When a session settles, Mend harvests the harness's own state into the store, so mend resume restores it natively (a Claude session resumes with its memory intact, even across machines), and --with carries the conversation into a different harness.

The reviewable object is the session's accumulated local change — worktree versus base — with the session record beside it. Review it in the terminal dashboard or the Mend web app. No issue tracker or pull request required.