orchia
v0.2.0
Published
One local dashboard for all your Claude Code sessions, organized by project
Maintainers
Readme
Orchia
Run five Claude Code sessions at once without babysitting five terminals.
Orchia is a local dashboard for all your Claude Code sessions, organized by project. See at a glance which ones are working, which ones need you and which ones are done, and allow or deny their permission requests straight from the grid.
npx orchiaThen open http://127.0.0.1:4317. Orchia starts the real claude CLI in terminals it hosts
and listens to it through Claude Code hooks: your settings, skills and subscription stay as
they are, and nothing leaves your machine.
25 seconds. Click for the video with sound.
Why Orchia
- One inbox for every session. Sessions waiting on a permission, a question or a plan come first, with Allow and Deny right there. No more tabbing through terminals to find the one that has been stuck for ten minutes.
- Plans in lots. When Claude proposes a multi-step plan, split it: lot 1 starts now, the
rest waits in Up next and is sent one at a time as Claude finishes, with a
/compactin between when the context fills up. - GitHub issues to commits. Pick open issues: each becomes a queued task that ends in a
commit with
Fixes #N. - Its own copy of the repo. Let a session work in an isolated copy, then review and merge it.
- Security alerts. Your connectors, skills and plugins are checked against known CVEs.
- Local and private. Listens on
127.0.0.1only. No telemetry, no account.
A look around
The demo and the screenshots show the Établi theme with made-up projects, rendered from
docs/design/mockups.html. Open it
in a browser to page through the screens, or add #grid4-dark (any screen id and
light or dark) to the URL to render one screen at full size.
Requirements
- Node.js 20.19+ or 22.12+
- Claude Code installed and signed in
(
claudemust work in a terminal) - Git, for the inspector's changes view and for sessions in their own copy of the repo
- Optional: the GitHub CLI, signed in, to queue GitHub issues
Developed and used daily on Windows 11. CI builds on Windows, macOS and Linux, but macOS and Linux have had little real use yet: reports are welcome.
Run
npx orchia # or: npm install -g orchia, then orchiaFrom source:
git clone https://github.com/MPoulhazan/orchia.git
cd orchia
npm install
npm start # builds the UI, serves everything on http://127.0.0.1:4317
npm run dev # development: UI with hot reload on http://localhost:5173Reloading the page reattaches to running sessions. Projects are saved in
~/.orchia/config.json and sessions in ~/.orchia/sessions.json.
Orchia was called AgentHub until October 2026: on first start it moves ~/.agenthub to
~/.orchia, so nothing is lost.
When the server stops (restart, closed terminal, reboot, crash), sessions come
back paused and resume their Claude conversation (claude --resume) as soon
as they are shown. A turn that was running at that moment is interrupted. The
restart button keeps the conversation; "New conversation" starts a fresh one.
Shortcuts
| Keys | Action |
| ------------- | -------------------------------------- |
| Ctrl K | Search projects, folders and actions |
| Alt N | New session in the current project |
| Alt G | Switch between focus and grid view |
| Alt I | Show or hide the inspector (focus view) |
| Alt 1 … 9 | Jump to a session (sidebar order) |
| Shift Enter | New line in Claude's prompt |
| Ctrl C | Copy when text is selected, else interrupt |
A new session is named after your first prompt, then after the title Claude Code gives
the conversation (the one /resume lists). Double-click a session in the sidebar to
rename it: a name you type is kept. In grid view, clicking a
session in the sidebar adds it to the grid (up to 6); double-click a cell header
to focus it. Alt 1…9 uses the physical number keys, so it works on AZERTY too.
Session status
Sessions started from Orchia report what they are doing through Claude Code
HTTP hooks, injected with --settings ~/.orchia/claude-hooks.json. They are
added next to your own hooks, never replacing them, and only for these sessions.
| Dot | Meaning | | ------------ | ---------------------------------------------------- | | green, pulse | Working | | amber | Needs you: permission, question or plan to review | | blue | Done, not looked at yet | | gray | Idle | | hollow | Paused after a server restart, resumes when shown | | dim | Process ended |
Amber and blue sessions are listed under Needs you and counted in the tab title. The bell in the sidebar turns on desktop notifications, sent only while the tab is in the background.
Permissions
When Claude asks to run a command, edit a file or use another tool, the request shows in Needs you and at the bottom of its grid cell, with Allow and Deny. Claude Code's own prompt stays in the terminal: answer either one, and the first answer applies.
- Allow allows this one call, like "Yes" in the terminal. To allow a command for good, answer in the terminal.
- Deny refuses the call and stops Claude's turn, like "No" in the terminal. Type in the session to tell Claude what to do instead.
- Questions and plans to review need the conversation's context, so they only open the session.
This uses the PermissionRequest hook: Orchia keeps it open until you answer,
for up to 10 minutes. Deny rules in your Claude settings still apply.
Plans in lots
When Claude presents a plan made of numbered steps or "Lot 1, Lot 2…" sections, a bar under the terminal offers Split into Up next. You can edit or leave out lots, then Start lot 1: the plan is approved with lot 1 as the whole task, and the other lots wait in Up next, sent one at a time as Claude finishes. Approving in the terminal still approves the whole plan, as usual. The plan is cut along its own structure, without a model call.
Each lot asks Claude to commit its work and end with LOT n DONE, or LOT n BLOCKED: and
the reason. The next lot only goes after LOT n DONE: when a lot fails, ends on a question
or says neither, the rest waits and Up next says why, with Resume and Drop the rest.
So you can start a plan and walk away without coming back to lots piled on a broken one.
Outside plan mode too: when one of Claude's answers is a plan with titled lots ("Lot 1",
"Phase 2", "Task 3", or codes like "T1, T2, T3" in order, as headings or bold lines), the
same bar offers Run as a sequence. A plain numbered list doesn't count, so a summary of
what was done doesn't trigger it. Whatever shape Claude's plan took (a table, "T1, T2"…),
type /lots in the session (or Run the last plan as lots in Ctrl K): Claude restates
it as ### Lot N: lots and the lot editor opens right away. /lots comes from a small
plugin Orchia loads with --plugin-dir in its own sessions only; claude run outside Orchia
is unchanged.
Before starting, pick what happens Between lots: Compact when needed (the default,
see below) keeps the conversation, Clear gives each lot a fresh one. With Clear, Orchia
types /clear before each lot after the first, and that lot's prompt carries the whole plan
and points Claude to git log for what is done, so Clear always commits after each lot.
Commit after each lot can be turned off otherwise.
While the lots run, Up next shows which one Claude is on ("Lot 2/4"). Desktop notifications and "Needs you on your phone" stay quiet between lots: you hear about the plan once, when all its lots are done or when one is held.
Between two Up next prompts, when the conversation fills more than 60% of the context,
Orchia types /compact first, so Claude Code doesn't compact in the middle of the next
task. Turn it off in Settings, under Compact between tasks.
GitHub issues
From an issue in Up next (or Work on GitHub issues… in Ctrl K) lists the
repository's open issues. Check a few and press Enter: each becomes an Up next prompt
that has Claude read the issue with gh issue view, fix it, and commit with
"Fixes #N", so GitHub closes the issue when you push. For a big issue, tick Plan
first: Claude plans it in lots, which you split into Up next as above.
Issues opened by someone who is not an owner, member or collaborator are hidden until
you ask: their text goes to Claude, so read it in the preview first. This needs the
GitHub CLI signed in (gh auth login); Orchia uses your gh
login and stores no token.
Model and effort
The chip next to the session name shows the model that last answered (read from
the transcript and the PostModelSwitch hook) and the effort chosen in Orchia.
Changing them relaunches that session with --resume --model … --effort …: the
conversation is kept and your global Claude settings are not touched (unlike
typing /model, which saves a new default). An /effort typed inside Claude is
not reflected in the chip.
Inspector
The panel on the right of the focus view (Alt I) shows, for the session on
screen:
- From the last answer: the code blocks of Claude's last reply, each with Copy.
A block that is one shell command also has Run, which types
! commandin Claude's prompt: Claude Code runs it in the session's folder and Claude sees the output. Run waits until Claude is waiting for you. - Context used and cost so far, as Claude Code itself estimates them. Claude
Code only gives these to its status line, so the injected settings also set a
status line command,
server/statusline.mjs, which passes them to Orchia and then runs your own status line (from your user or project settings) unchanged. Side effect: with any status line set, Claude Code hides some footer hints such as "esc to interrupt". - Skills and Connectors of the project: turn them on or off, add, sign in, remove. Connectors load when Claude starts, so after a change Orchia restarts the project's sessions by itself once Claude is free, keeping the conversation (Settings: Restart by itself).
- Plan, when Claude keeps one with its task tools (
TaskCreate/TaskUpdate, orTodoWrite). Recent models don't use them by default, so it is often absent. - Uncommitted changes and branch of the project folder, from
git statusandgit diff HEAD. Every session in the same folder sees the same changes.
Security alerts
Connectors, skills and plugins are checked against the advisories listed in
awesome-vetted-claude-code,
a catalog where each tool is reviewed and checked for CVEs. A shield next to an item in
the inspector links to its review: red when the issue applies (the version in your
command is affected, or there is no fix yet), grey when it was fixed in a version Orchia
can't see, so you can check yours. High and critical issues that apply also show in
Needs you. npx package and hosted servers run the latest release, so only issues
with no fix yet concern them.
Orchia also watches for secrets that reach the model: a token in your prompt, or in a
tool call's input or output (cat .env printing an API key). Only formats with a fixed
prefix are recognized (GitHub, Anthropic, OpenAI, AWS, Stripe live, GitLab, Slack, npm,
Google API keys and private keys), so an alert is almost always real. It shows in
Needs you as what happened ("Stripe live key sent to Claude"), where, and Revoke on
Stripe; the command or file it was seen in shows on hover. AWS's documented example
keys (…EXAMPLE) are not secrets and never raise an alert.
Orchia keeps only the kind of secret, its first and last characters (ghp_…a41f) and a
hash to recognize it again, in ~/.orchia/secrets.json; never the secret itself.
The same rules guard git: when Claude is about to run git commit or git push, Orchia
reads what would go in (the staged changes, plus what the command stages itself with
git add or commit -a; for a push, every commit it sends) and refuses the command if it
adds a secret. Claude is told which secret, masked, and in which file, so it takes it out
and commits again; the alert shows in Needs you and on Orchia Cloud. Only Claude's
commands are checked, not commits you type yourself.
Permission settings that let Claude do more than you may think show in Needs you too, for projects with a running session:
- an allow rule for every shell command (
Bash,Bash(*)), with Remove rule when it is in the project's.claude/settings.local.json; defaultMode: bypassPermissions, which skips every prompt;.envfiles in the project, subfolders included, with no deny rule: Hide from Claude addsRead(**/.env)andRead(**/.env.*)to the deny rules of.claude/settings.local.json. Templates (.env.example…) and files holding only browser-publicVITE_… values don't raise the alert, but the rules can't leave templates out, so Claude can't read those either. Deny rules don't see the shell, so once reads are blocked Orchia also refuses commands that print a.envfile (cat,head,grep,Get-Content…); loading one without printing it (source .env, dotenv) still works.
What a project already has in git is checked too, when the project is first opened, then once a day:
- a
.envfile that is tracked (templates such as.env.example, and files holding only browser-publicVITE_/NEXT_PUBLIC_… values, don't count); - a private key that is tracked (
id_rsa,*.p12, a.pemor.keyholding a private key); - a secret written in clear in a tracked file, with the same fixed-prefix rules as above
(AWS's documented
…EXAMPLEkeys are skipped); - a
.envfile git doesn't ignore, onegit add -Aaway from a commit: Add to .gitignore adds it; - GitHub's secret scanning or push protection turned off on a public repository you
administer, where both are free: Turn on enables them with your
ghlogin. They are a second guard, for commits made outside Claude too. Private repositories need GitHub's paid Secret Protection, so Orchia says nothing there.
Needs you only shows the alerts that can't wait: a secret that reached the model or a
commit, a serious CVE, permissions left wide open, a secret in git. The others (readable
.env files, a .env git would commit, GitHub's scanning off) can wait. Security · N
open, under the projects next to Plan usage, opens the full list: what Needs you shows,
what you put off, what can wait, and what was decided in the last 90 days, with Reopen.
✕ on an alert puts it off (Later): it leaves Needs you without a decision and waits in the Security list. A secret comes back as soon as it is seen again; anything else after 30 days. Orchia Cloud doesn't know about Later: there, the alert stays open until it is decided under Resolve:
- Fixed: on its own for a CVE or a setting, as soon as Orchia no longer sees the problem; I revoked it for a secret, which only you can do.
- Accept the risk… asks why, then hides the alert for 90 days. It comes back once after that, to be looked at again.
- Not a real secret (or Not a real issue) hides it for good: a test token, an example key.
Decisions are kept in ~/.orchia/triage.json, alerts put off in ~/.orchia/later.json.
Environment
| Variable | Default | Purpose |
| --------------- | ------------ | ------------------------------- |
| ORCHIA_PORT | 4317 | Server port |
| ORCHIA_HOME | ~/.orchia | Where projects are saved |
| ORCHIA_CLAUDE | claude | Command used to start Claude |
| ORCHIA_CLOUD_URL | https://app.orchia.app | Orchia Cloud server, if you pair |
Security
Orchia starts processes on your machine, so it is built to be reachable only by you:
- The server listens on
127.0.0.1only, never on your network. - It rejects requests whose
Hostis notlocalhostor127.0.0.1(blocks DNS rebinding) and whoseOriginis another site (blocks cross-site requests from pages open in your browser). - Each session gets its own random token; hook calls without it are refused.
- Nothing leaves your machine except what Claude Code itself sends. There is no
telemetry. Orchia itself only downloads the public advisory catalog
(
tools.jsonfrom awesome-vetted-claude-code), once a day, without sending anything about you or your projects. - Orchia Cloud is optional and off by default. If you turn it on in Settings, Orchia
pairs with your account through a code you approve in the browser, then keeps one
outbound connection open (the cloud never connects to your machine). It sends this
computer's name, Orchia's version, your OS and when you are online, nothing else:
no code, prompts, terminal output, file or project names. Settings lists exactly what
is sent. The device token is saved in
~/.orchia/cloud.json; Disconnect, or Unpair from Devices on the cloud, revokes it. SetORCHIA_CLOUD_URLto use another server. - "Needs you on your phone", a second switch that is also off by default, sends the sessions that need you to the cloud so you can allow or deny from your phone: each one's session and project name, why it waits, a permission request's tool and target (the command, file or URL), and the summary of a finished turn. Still no code, prompts or terminal output. The phone can only answer a request Claude already made, never start one or type in a session, and "Always allow" stays on your computer. Settings can limit it to Deny, or keep shell commands for the computer. Claude Code's own deny rules still apply.
- "Security alerts in Orchia Cloud", a third switch, also off by default, lists this
computer's open security issues on the cloud's Security page: CVEs in your connectors,
secrets that reached the model and risky permission settings, each with its severity,
a short description, the project and session names, and what you decided about it (an
accepted risk with your note, or a false positive). A secret is only named by its
kind and first and last characters (
ghp_…a41f), never its value. Turning the switch off removes this computer's issues from the cloud. - A connector token you paste is written to Claude Code's own config
(
~/.claude.json), whereclaude mcp addwould put it.
Anyone who can run code as your user can already drive claude, so Orchia does not
try to protect against that. If you find a security issue, please open a
private security advisory
rather than a public issue.
Contributing
Bug reports, ideas and pull requests are welcome. See CONTRIBUTING.md.
License
MIT. Orchia is an independent project, not affiliated with or endorsed by Anthropic. Claude and Claude Code are trademarks of Anthropic.
