dsh-oh-my-claude

Oh My Claude

English | 中文

Claude Code, native inside dsh

Drive your logged-in Claude Code CLI as a first-class dsh provider: no API key, no extra services, no runtime dependencies.

Built with AI assistance, and open to more. Contributions from AI coding agents are welcome; start with CONTRIBUTING.md.

Claude Code CLI as an LLM provider for dsh. Every request drives claude -p with stream-json in and out, so it uses whatever login, hooks, CLAUDE.md files, MCP servers and rate limits Claude Code already has. No API key needed. The plugin speaks the CLI's own protocol (the one the Agent SDK wraps) directly, so it has no runtime dependencies.

Tour

The shield is dsh's own control. In a Claude session its rows become Claude's six permission modes, each labelled with the dsh access level it sets underneath.

One ✻ button beside the composer opens the whole plugin: Memory, Instructions, Skills, Rewind, Changes, MCP, Asides, Diagnostics, Tasks and Tune, plus a Restore tab while the session is still blank.

Cost and cached token count, two figures dsh cannot compute, join dsh's footer stats row. The pill turns orange once a session passes the spend line you set. The dollar figures are what the turns would cost at API rates; on a Claude subscription you are not billed them.

Plan usage and the CLI's own context breakdown live in dsh's context ring popover. On a phone the panel becomes a sheet.

Once an SSH box is saved, dsh's own Add workspace dialog gains a box dropdown: a folder on that box becomes a workspace here, and the session in it runs Claude Code there.

Shots by tools/playwright/tour.ts.

Install

You need the Claude Code CLI on PATH and logged in (claude --version works, claude opens without asking you to sign in), and dsh 0.2 (0.2.0-rc.1 or newer). On dsh 0.1.x, install the npm tag for your line from the Compatibility table below. Nothing else: no API key, no Node build step.

From dsh's Plugins page, in the web app or DSH Desktop:

1. Open Plugins in the sidebar and press Add plugin.
2. Enter dsh-oh-my-claude and press Install.
3. Press Enable now, or turn on its switch later under Installed plugins.

It loads without a restart. Reload the page and "Oh My Claude" appears in the model picker with the models your login can use. Pick one and chat.

From a terminal, then restart dsh however you run it:

dsh plugin --profile web add dsh-oh-my-claude
systemctl --user restart dsh-web.service

To install straight from the repository instead of npm, use github:lcestou/dsh-oh-my-claude as the package name in either route.

The plugin speaks English and Chinese and follows dsh's own language setting (Settings → General → Language), switching as soon as you change it.

If the picker lists it as Oh My Claude (not logged in), run claude auth login in a terminal on the box that runs dsh, or press Log in under Settings → Oh My Claude → Boxes.

DSH Desktop (macOS and Windows) uses the Plugins page steps above. The plugin looks for claude in the usual install folders, since an app started from the Dock does not see your shell's PATH. On Windows it runs Claude as a plain child of dsh, so restarting the app ends a running turn, and the parts that need a Unix shell (Log in from Settings, SSH boxes) are not available. The server half has been run under Electron on Linux the way Desktop runs it; a real Mac or Windows install has not been tried yet.

When a newer version is on npm, an orange pill with the version number appears in Settings and on the panel's Runtime line; a click copies the update command for a terminal. On the Plugins page, dsh updates a third-party plugin by removing it and adding it again. Details, and the repairs the plugin makes on its own after a dsh upgrade, are under Plugin updates.

Optional, in ~/.dsh/settings.yaml:

agent-default-model:            # make Claude Code the default for new sessions
  provider: claude-code
  model: claude-fable-5-1       # any model id the picker lists
subagent-model-selection:       # let dsh subagents run on Claude Code too
  enabled: true
  allowedModels:
    - provider: claude-code
      model: claude-haiku-5-5   # any model id the picker lists

Working on the plugin itself? Start at docs/developing.md.

Compatibility

dsh ships prereleases only and moves client APIs between them, so every release of the plugin records the dsh line it was written for and the dsh it was driven on. "Runs on" is the range the code carries branches for; "tested on" is the dsh installed on the box when the release was cut. The update pill reads the floor off the registry and never offers a release to a box whose dsh is below it.

Plugin · Runs on dsh · Tested on dsh · npm tag 1.5.7 · 0.2.0-rc.1 through 0.2.1-rc.3 · 0.2.1-rc.3 · latest, dsh-0.2 1.5.6 · 0.2.0-rc.1 through 0.2.1-rc.3 · 0.2.1-rc.3 · none, superseded by 1.5.7 1.5.5 · 0.2.0-rc.1 through 0.2.1-rc.2 · 0.2.1-rc.2 · none, superseded by 1.5.6 1.5.4 · 0.2.0-rc.1 through 0.2.1-rc.2 · 0.2.1-rc.2 · none, superseded by 1.5.5 1.5.3 · 0.2.0-rc.1 through 0.2.1-alpha.1 · 0.2.1-alpha.1 · none, superseded by 1.5.4 1.5.2 · 0.2.0-rc.1 through 0.2.0-rc.2 · 0.2.0-rc.2 · none, superseded by 1.5.3 1.4.2 · 0.1.5-rc.1 through 0.1.7-rc.2 · 0.1.7-rc.2 · dsh-0.1.7 1.4.1 · 0.1.5-rc.1 through 0.1.7-rc.2 · 0.1.7-rc.2 · none, superseded by 1.4.2 1.4.0 · 0.1.5-rc.1 through 0.1.7-rc.2 · 0.1.7-rc.2 · none, superseded by 1.4.1 1.3.1 · 0.1.5-rc.1 through 0.1.6-alpha.2 · 0.1.6-alpha.2 · dsh-0.1.6 1.2.1 · 0.1.5-rc.1 through 0.1.6-alpha.2 · 0.1.6-alpha.2 · dsh-0.1.5

What you get

1. Any Claude model your login can use, in dsh's own picker. Models
2. Every dsh session resumes its Claude Code session, tool history included, and a terminal claude /resume sees the same one. Sessions
3. Restart dsh without killing a running Claude turn. Restarts
4. dsh's access shield sets Claude's six permission modes per session. Permission mode
5. Claude's permission prompts, questions and plan reviews become dsh dialogs, each option's preview under the question. Approvals
6. One spark button opens Memory, Rewind, Changes, MCP and seven more tabs, all reachable from the keyboard. The panel
7. Cost and cached tokens in dsh's footer, with a spend line that turns the pill orange. Cost pill
8. Plan usage (5-hour, weekly, extra usage) in dsh's context ring, which turns amber near a limit and red at one, with a notice above the composer. Only limits the session's model counts against, graded by Anthropic, so a new model needs no update. Plan usage
9. Restore or import any past Claude transcript, from dsh or from a terminal, and search inside them for what was said. Session browser
10. Claude's slash commands and skills show up in dsh's slash menu as /claude-. Command bridge
11. Claude gets dsh's subagents, jobs, goals and web search over MCP. dsh tools over MCP
12. Edit CLAUDE.md, settings.json and MCP servers from the panel, on this box or a remote one, and sign in to an MCP server that needs it. Instructions, settings.json editor, MCP
13. Run Claude on another box over SSH, or mount several accounts. Remote boxes and accounts
14. Report a problem with a redacted box report, copied or filed as an issue. Report a problem
15. All the orange is a switch, per group, any colour. Claude look
16. Pick what dsh adds to your prompts, with the size of each block on the label. dsh context
17. A card offers each new Claude Code release for the box a session runs on, and installs it from dsh; a headless claude -p never updates itself. Channel, auto-install and history per box sit under the switch in Settings. Claude Code updates
18. What the last five versions changed, read from the plugin's own changelog, in Settings. Changelog
19. Every skill the CLI can reach, grouped by scope, with add, edit and remove for your own and the project's. What each costs, from the CLI's own /skill-doctor, is a table you can sort, beside what else drives your usage: the MCP servers behind the most requests, and how the work was shaped. Skills
20. A mark on the browser tab and an optional desktop notice when a session you are not looking at finishes, or stops to ask you something. Session notices
21. dsh's own goal and every earlier goal of the session, beside Claude Code's goal and its scheduled tasks. Tasks
22. Edit or remove a message you steered in while Claude works, or all of them as one, until Claude reads it. A file or image on the message stays on it; only the text changes. Nothing goes out while you edit. Send now stops Claude and sends one as the next turn. Steering
23. Claude's working line (spinner, verb, figures) repeats above the composer once you scroll up past it, so reading back through a long turn never hides what Claude is doing. A switch in Settings turns it off. Turn status
24. Select text in a reply to quote it into your message, or to ask Claude about it as a side question that leaves the conversation alone; leave the question blank and Claude explains the passage. A quote in your own message reads like one in Claude's replies, indented with the accent bar. Asides
25. A finished turn keeps a closing line where the working verb was, the way the terminal ends one: ✻ Crunched for 27s · done 3:21 AM. A turn you stopped reads Interrupted · What should Claude do instead?. Turn status
26. On dsh 0.1.7, Claude's tool calls show as dsh's own tool cards, the way every other provider's do, with the text streaming between them. Inline text is a switch away in Settings. Tool rows
27. A new session in a workspace opens on the Claude box and model that workspace last ran. Remember model per workspace
28. After each reply, Claude's guess at your next prompt shows above the composer: Use writes it into the box, Send sends it as it is, and it folds away while you type something else. On by default, as in the terminal; a switch in Settings turns it off. Suggested next prompt
29. Claude's background shell tasks and subagents are listed on the Tasks tab while they run, each with Stop and, for a shell task, Output. A subagent's row in the chat names each step as the CLI reports it. Tasks
30. If an ANTHROPIC_API_KEY in dsh's environment is what Claude Code bills instead of your login, the session says so in one line at its start. Where things live
31. Each plan window says where it is heading at the current rate: the percent it ends at, or when the cap lands. Plan usage
32. Once the context window is 75% full (adjustable in Settings), a card above the composer says so and offers Compact, which sends Claude's own /compact. Context nudge
33. A Run button on every shell snippet Claude writes puts it in the composer as a prompt; Enter sends it, and it runs through Claude's Bash tool and dsh's approval. Run a snippet
34. A session renamed in dsh keeps that name in claude --resume, and a name set on the Claude side renames the dsh session. How it works

Defaults you never have to touch, such as secret redaction in tool results and surviving the CLI's own updates, are on How it works.

Configure

Settings come in two kinds. The switches under Settings → Oh My Claude and in the panel's Tune tab apply at once and are kept on the box, so every browser sees the same ones. The keys below are install-time: they go as config: on the bundle row in the profile's cordis.patch.yml and take effect when dsh next loads the plugin, which a dsh restart does. dsh's own Plugins page lists the plugin with an on/off switch but has no form for these keys. Three keys people change first:

permissionMode: dsh      # follow the session's shield (default), or pin one of the CLI's six modes
spawn: keeper            # Claude outlives a dsh restart (default); node or dsh for a plain child
sshHost: ""              # "[user@]host" drives a claude on another machine over ssh

Every key, with its default, is in docs/configuration.md.

Remote boxes and accounts

Mount the plugin twice with different configDir values for two logins. List other boxes running dsh and this plugin under Settings → Oh My Claude → Boxes and jump between them. Or give a box a name and user@host and drive the claude there over SSH, with nothing on the far side but the CLI; Tailscale and WireGuard peers appear as hosts to pick from. All of it is on docs/remote.md.

What it touches

• Runs the claude binary you already have, on this box or on one you add over SSH.
• Reads Claude Code's login (~/.claude/.credentials.json, or ANTHROPIC_API_KEY when set) to fetch the model list and your plan usage. That token goes only to api.anthropic.com.
• Writes its own state under ~/.local/state/dsh-oh-my-claude/; login tokens for SSH boxes are files only you can read. CLAUDE.md, settings.json and MCP servers change only when you edit them from the panel.
• Listens on a Unix socket for the keeper that holds Claude across a dsh restart. Its HTTP routes live inside dsh's own server, so on DSH Web it opens no TCP port. Under DSH Desktop, whose web server answers 403 to anything but its own window, the bridge that hands dsh's tools to Claude also listens on 127.0.0.1, behind the same key.
• Calls out to api.anthropic.com (models, plan usage), status.anthropic.com, downloads.claude.ai (the CLI's release pointer), registry.npmjs.org (plugin update check) and api.github.com (the repo's star count). SSH goes only to boxes you add.

Not covered

• dsh's shell and file tools are not proxied (except bash for background jobs); Claude Code uses its own, under its own permission mode.
• Claude Code sessions are not deleted when dsh sessions are.

Bugs and feedback

Found a bug, or something that reads wrong? Open an issue at github.com/lcestou/dsh-oh-my-claude/issues with the dsh and Claude Code versions (dsh --version, claude --version) and what you expected to see. Settings → Oh My Claude → Report a problem writes that report for you.

License

MIT.

Claude, Claude Code and the Claude spark mark are trademarks of Anthropic, PBC. This project is an independent dsh plugin that drives the Claude Code CLI you already have; it is not made, endorsed or supported by Anthropic. The spark is drawn here only to say which sessions and controls belong to Claude Code, in the way the CLI itself draws it.