Partitions
A separate world for every agent.
The partition is the unit everything in Ahura OS lives in. It is how state stays isolated, how cost stays attributable, and how you pause a project for a week or a century.
An agent can create as many partitions as it likes. There is no cap on how many you create.
Providing data
Two ways to work with your data.
Every partition is isolated. What differs is whether your agent is allowed to see what is inside. Choose per partition, depending on how much you trust the agent with the data.
Agents work with the data directly
Regular partition
The default. Your agent can read, write and search everything in the partition, and can send data to other systems as the work requires. Use it when you trust the agent with the data, which is most work.
- • Agents read files and records straight into their context
- • An agent can work across your regular partitions by naming the partition on each call, so it can read from one and write to another
- • Every Ahura OS tool is available
Agents direct the work but never see the data
Privacy partition
Create a partition as Privacy and no data ever leaves it, not to other partitions, not to outside systems and not into an agent’s context. The agent instructs Ahura OS to read, parse, index, search or process, the work runs inside the partition, and the agent gets back results and references only.
- • Built for zero-trust agentic workers
- • Set when the partition is created, from the UI or by an agent, and it cannot be changed afterwards
- • Tools that would return raw content to the agent are blocked or limited to references
- • Never a source or destination for data moving to or from another partition
- • Good for contracts, medical records and customer data
In a Privacy partition, Ahura GPU, Ahura Search and Ahura Store all work on data in place, so heavy processing never needs the data to travel.
Uploading
Upload data to any partition.
You do not need an agent to get files in. Add them yourself from the UI, whichever mode the partition is in.
Any partition
Upload files from the Ahura OS interface into a regular or a Privacy partition.
No agent in the path
Files go from your browser into the partition. Your agent never handles the bytes, so its context stays free of them.
Ready to use
Uploaded files sit in Ahura Store inside the partition, and agents can then ask for them to be read, indexed or processed.
In a Privacy partition this is how data arrives at all: you upload it, and the agent only ever instructs.
Create as many as you need
Partitions are cheap to make and easy to throw away, so use them freely. Agents create them with a single call, with no approval step and no quota to request.
One per task
Give every job a clean workspace and delete it when the job is done.
One per customer
Keep each customer’s data apart without any extra design.
One per experiment
Try something risky, keep the results, discard the rest.
Short-lived scratch space
Set a lifetime and the partition destroys itself when it expires.
Each partition is billed for what it stores and runs, and your spending limits still apply, so unused partitions are worth destroying or freezing.
- 01
Create
One call makes a partition. Every service your agent uses runs inside it, so memory, files, records and workflows stay together and stay isolated from everyone else.
- 02
Live
Read, write and run as normal. Usage is metered against your credits as it happens.
- 03
Freeze
When work pauses, freeze the partition. Its storage moves to low-cost archive pricing and compute is released. A manifest records exactly what was there.
- 04
Rehydrate
Thaw it and the manifest restores everything to the state it was frozen in.
- 05
Destroy
When a partition is finished, destroy it. Anything still owed back to you is returned.
Every call, listed
13 tools your agent can call. Each one is a standard MCP tool with a typed schema and a price shown in tools/list.
Create and inspect
- partition.create
- Make a partition with an optional lifetime, retention window and, for a Privacy partition, a result policy. Anyone can create as many as they need.
- partition.get / partition.list
- Status, storage used live and frozen, item and object counts, hourly rate, expiry, retention and any running job.
- partition.jobs.get
- Progress and errors for a freeze or restore job.
Lifecycle
- partition.freeze
- Move a partition to low-cost archive and release compute. Runs as a background job with a manifest of what was there.
- partition.rehydrate
- Restore a frozen partition from its manifest, verified before anything is brought back.
- partition.extend
- Push out the expiry or the retention window.
- partition.destroy
- Remove a partition in any state, including frozen, and return what is owed.
Privacy and access
- partition.roles.get
- The roles on a partition. An agent sees its own role.
- privacy.policy.get
- The result policy and release mode of a Privacy partition.
- privacy.audit.list
- Metadata-only log of every instruction and release in a Privacy partition.
- privacy.release.get
- Check whether a held result has been approved.
Identity
- identity.partition.get
- The partition’s public key and fingerprint.
- identity.partition.sign
- Sign with the partition’s key, for manifests and attestations from this partition.
Some controls are deliberately not agent tools. Uploading files, changing roles, changing a Privacy partition’s policy and approving held results are done by a signed-in person in the UI, so an agent cannot do them for itself.
Isolated
Each partition is its own world. Data moves between regular partitions only when an agent does it, and never into or out of a Privacy partition.
Time-limited
Set a lifetime and a retention window. Expired partitions clean themselves up.
Long-term
Frozen state can be kept for decades at archive cost.
Predictable
Credits are consumed per partition, so each project has its own bill.
Private on request
Create a partition as Privacy and no data leaves it, and agents direct the work without seeing the data.
Freezing, in plain terms
Two things can freeze a partition: you, or an empty balance. Running out of credits locks the partition read-only instead of deleting it. A seven-day grace period follows, so you can top up and carry on. Only after that does teardown begin, and frozen state is archived first.
Who uses it
- DAOs and protocols keeping governance agents' memory on ice between cycles.
- Research teams preserving experiment state long after a project ends.
- Product teams pausing per-customer agents without paying for idle servers.