No context window selected
Create or import a context window to begin.
Settings
These settings apply to
Account
Product tour
Review context windows, models, imports, tools, collaboration, and privacy. Your place is saved if you leave before the end.
Name & email
PNG, JPEG, or WebP · up to 1 MB
Changing the email requires your current password — it is how you sign in and where reset links bind.
Change password
At least 8 characters. Changing it signs out every other device.
Language & translation
Choose the language you prefer to read. ContextCloud offers translation only when language detection reaches 72% confidence. Translation changes the display only: the original message remains canonical for quotes, search, exports, and audit history.
When another language is detected
Your Spaces
Everything you work on lives in a space. Personal spaces are private to you; group spaces are shared. Pages that can vary by space show a scope selector above their settings.
Security
A second factor protects your stored logins and admin actions even if this browser session is stolen.
Your sessions
Devices and apps signed in to your account. Revoke anything you do not recognize; the revoked device is signed out immediately.
Notifications
ContextCloud always keeps in-app notifications under the bell. Browser alerts are optional and are controlled by this browser.
Browser alerts
Organization
Everyone with an account here belongs to this organization. Connected apps and skills you set for the organization reach every member automatically, and members cannot remove them. Provider keys work differently: the organization's key backs group spaces, but not members' personal spaces unless you allow that under AI Model Access. Where a member or their group has its own key or default model, that one is used instead.
Reopens the administrator setup and product tour — organization name, invitations, a company AI key, windows, models, imports, tools, collaboration, and privacy. Anything you leave blank stays as it is. Email addresses you enter are invited as you move through the steps, and a key entered there is saved to this organization under its company-key policy.
Personal context windows
These organization-wide rules apply immediately. Existing personal windows keep their creation-time owner, even if you change the default for new ones.
Organization-owned personal windows remain recoverable by organization admins if the member leaves. Individually owned windows are permanently deleted when their owner is purged.
Membership
When this is off, group owners can still change or remove current members, but every add and invite action is hidden and blocked. Organization admins can always manage membership.
Cost estimates
Cost figures shown to members are estimates built from published list prices. If your organization pays less than list, work your share out once: divide a real bill by the estimate ContextCloud showed for the same period — a $45 bill against a $283 estimate gives 0.16 — and enter that. Leave it at 1 to show list prices unchanged. Members see the scaled figure, and spending limits are measured against it, so a number below 1 lets people run more before a limit stops them. Values are kept between 0.01 and 10.
Audit log
Investigation history for sensitive actions across this organization. Records are centrally redacted, linked into a tamper-evident chain, and visible only to organization admins. Open a row for bounded details; secret values and full prompts are never shown.
Audit details
Organization defaults
Model access, the default model, and tools every member inherits in their personal space.
Members
Desktop App
The desktop app runs ContextCloud alongside your computer so it can browse and import your local Claude and Codex sessions, open artifacts, and connect local tools — things a browser tab cannot do.
Usage and cost
What has been spent on AI this month. Every dollar here is an estimate, not a bill.
Spending over time
Estimated spending across your whole organization, last six months. Only org admins can see this.
Most expensive context windows
One row per piece of work this month, for the whole organization — not just the scope selected above.
Spending limits
A finite organization-wide limit protects paid AI work by default. Group admins can add tighter limits for groups they manage.
Server
System-wide settings only server admins can change. They apply to every organization member.
Organizations
Each organization is a fully isolated tenant with its own members, windows, credentials, and data. Create one, then share the sign-in link with its admin.
Release
Running release
Reading the running release…
Opens the protected promotion action. The release lane still verifies the candidate, signed evidence, approvals, and production preflight before changing production.
Operations
Live workload
Active and queued runs across every organization — whose run, which window, how long, and roughly what cost, with subagent, browser, and model signals. Prompts, transcripts, credentials, and artifact content never appear here; raw ids live behind each row's Details.
Loading the live workload…
Organizations & limits
Live use against caps. Workspace defaults apply to every organization without an override; run caps queue or refuse new runs per the workspace policy, and storage caps refuse new growth — nothing is ever evicted.
Loading organization limits…
Capacity & storage
Database volume capacity, growth forecast, reserve state, alert health, and per-organization storage attribution. Automatic cleanup stays limited to rebuildable cache and temporary data — originals and sole restore points are never eviction candidates.
Loading the capacity dashboard…
Authentication & notifications
Transactional email
Used for forgot-password links, organization invitations, and other security-critical messages. Provider credentials stay in the server deployment, never in an organization or user account. Checking…
Model pricing & keys
Open-source model rate
Estimate used only for open-source models that are not in the catalog — every catalogued model bills at its own published rate. Leave blank to use the built-in estimate (3 in / 15 out USD per million tokens, deliberately pessimistic). Setting a value here applies that one flat number to input and output alike.
Web search key (Brave)
One Brave Search API key for the whole server; web_search uses it so research runs don't depend on scraping (which cloud IPs get blocked from). Get a key at
brave.com/search/api. Stored encrypted at rest. Not setImports
Post-import processing
What happens to a conversation right after it is imported. The digest lists the session's deliverables and website preview; the AI summary additionally queues one small background run that writes a handoff paragraph into the digest card (needs a configured provider key — skipped silently otherwise).
Backups & storage
Backups
Daily archive of everything durable — transcripts, artifacts, memories, settings, usage — excluding scratch workspaces. Download a copy to keep one off the server; restore with
scripts/restore-backup.sh.Offsite backup storage
Mirror each backup to your own object storage so a copy survives if this server's disk is lost. Enter your account keys — they are encrypted at rest. Not configured
App & disk
Desktop app download
Upload the signed macOS installer (.dmg). It is stored in the object-storage bucket configured above, and every signed-in user gets a Download button in their account menu. No build uploaded
Upload the build that opens to this server.
Disk & workspace retention
Measuring disk usage...
Scratch workspaces of context windows idle for more than this many days are cleared automatically (0 = off). Transcripts, artifacts, and memories are never touched; a notice appears in the window.
Storage & Memory
Every context window, how much room it takes on the server, and the notes it carries forward.
Workspace files are temporary files the assistant made while working — clearing them frees space and does not remove anything you can open. Artifacts are the finished documents it produced. Memories are notes brought over by a Claude import.
Members & Groups
The people in this organization — active, deactivated, and invited — and the groups they belong to.
Put people in a group and they can all open the same context windows, and the apps and sign-ins added to the group are available for them to connect in those windows. Anything outside a group stays with the person who made it.
Which model runs by default
The provider and model new runs start with in this scope. People can still pick a different model per run.
Skills
A skill is a set of instructions you write once — how invoices get handled, how a proposal is worded, how files are named — and the assistant follows them on its own instead of asking you again. A new skill does nothing until you switch it on. Skills set up for your whole organization, or by whoever runs this server, also apply to your tasks without appearing in this list.
Write a skill
Write the instructions once. Switch the skill on and the assistant follows them on its own every time a task matches the words you list below.
That reads like a sentence rather than a list of words. It still works — the assistant picks out the distinctive words, and it can load a listed skill by name when it judges the skill relevant — but short words separated by commas match most reliably.
Ordinary skill text is visible only to administrators of its owner. A secret skill is write-only even for them.
What a skill file can contain
Have the instructions in a document already? Upload it. A plain text file (.txt, .md or .markdown, up to 256 KB) works as it is — the file name becomes the skill name and everything in the file becomes the instructions. Uploading does not switch the skill on; it appears in the list below and you choose Enable when you want it used. If the file opens with a line of three dashes, the lines above the next line of three dashes can set name, description and triggers (also accepted: keywords, when, trigger_conditions), written as simple "key: value" pairs or a dash list. Nested or multi-line values are not read.Skills already added
Switch a skill on to let the assistant use it. A skill that is switched off stays here and changes nothing.
Waiting for approval
Credentials
Credentials are reusable sign-ins owned by the person, group, or organization shown in Settings. Manage the resource here; authorize it for a context window from that window’s Access screen.
Access for every window in this group
MCP servers
MCP servers and connected accounts are reusable resources owned by the person, group, or organization shown in Settings. Manage them here; authorize them for a context window from Access.
Action limits Loading current limits… Defaults
What the assistant is allowed to do
These limits apply to every context window in this scope. Set the most the assistant may ever do, then switch off individual actions app by app.
Any single context window can be set lower than this, never higher. Connected apps only work when this is set to "Everything, including connected apps" — the other two settings switch every connected app in this scope off. "Look things up only" still lets the assistant search and fetch web pages.
Actions, app by app
Switching an action off blocks it for every context window in this scope. While an allow-list is set under Advanced, only the actions named in it can run.
Kinds of action
Switch off a whole kind of action and it is blocked across every app and every context window in this scope, including apps added later. Untick "Export" and nothing here may download or extract data, whatever the app is called. Leave all five ticked to change nothing.
An app named on the allow-list under Advanced is still allowed, even if its kind of action is switched off here — naming one app is the more specific instruction.
Advanced — edit the action names directly
Type anything in the next box and everything not listed there is blocked — including the assistant's own abilities, like reading a file or searching the web. The only exception is the assistant's request for more tool calls, which always reaches you. Leave it empty unless you want that.
Saved entries are preserved whether or not they match an action we know about — but the switches above rewrite both lists from the last save, so anything typed here and not yet saved is lost the moment you flip a switch. Save here first, or use the switches, not both at once.
Connected apps
Apps already set up in this scope. "Built-in" ones come from ContextCloud; "Added here" ones were pasted in by someone in your workspace — check the address before you trust one.
Ready-made apps Browse connections ContextCloud already knows how to set up.
Ready-made connections
Choose Add, then sign in with your own account — nothing else to fill in. You can also just ask the assistant for one by name.
Access for every window in this group
Connect an app by hand
Use this when someone gave you an app's connection details to enter by hand. If the app appears in the one-step list above, add it there instead — that sets it up for you without any of these fields.
This takes two steps, and the first one usually needs whoever administers the other app. ContextCloud has to be registered with that app before the app will give out the client ID below. Show me how to do this
Step 1 — give this address to whoever registers ContextCloud. It goes in the box the other app labels Redirect URI or Callback URL:
Advanced sign-in settings
Fill these in only where the app's own instructions name a value. If you leave the permissions list empty, ContextCloud asks the app for it. The bottom two are only ever sent when you enter them — nothing is discovered for you. An app that runs a program on the ContextCloud server cannot be asked, so it always needs a sign-in address here.
Unsaved settings