A Letter of Marque for an Agent: Why the Fleet's New Feature Is Called Privateer
Sovereign Fleet lets coding agents work inside a kernel sandbox and lands their commit only when a gate is green. It is now learning to do work that is not code: read pages, pull numbers from an interface, write a report. The application keeps its name. The new feature needed one, and “bot” says nothing about what makes it different.
It is called Privateer.
The commission and the court
For centuries, until the Declaration of Paris ended the practice in 1856, a naval power did not build and crew every ship itself. It commissioned private ones. Two institutions made that something other than piracy.
The first was the letter of marque: a written commission from the sovereign that named what the ship was allowed to do, where, and for how long. A captain who sailed without one, or outside its terms, was a pirate under the law of every nation.
The second is the half people forget. A privateer did not own what it brought home on the captain’s word. It had to bring the prize into port, hand over the ship’s papers, and prove before a prize court that everything had happened within the commission. Only the court’s decision made the result count.
I am not interested in the plunder. I am interested in those two institutions, because they are the two things an autonomous agent needs and usually lacks: a permission that is written down and limited, and a judge who is not the captain.
What the name commits the software to
| In the story | In Sovereign Fleet |
|---|---|
| The sovereign | you, the operator: your machine, your keys, your data, the last word |
| The privateer | an agent with a third-party tool, sent out on one job |
| The letter of marque | the job’s commission: which tool, which domains, how many calls, how much memory, enforced by the kernel and a proxy |
| The ship’s log | a hash-chained record of every tool call and a record of every network request, both written where the agent cannot reach |
| The prize court | a gate that decides from those records and a schema, never from the agent’s own account |
The point of the picture is where each thing lives. In Sovereign Fleet, a letter of marque refers to a per-task capability contract enforced outside the model, because kernel filters and egress proxies hold even when a prompt fails. A prize court is defined as a verification gate that accepts or rejects from append-only call logs, response bodies, and egress logs rather than self-reports.
The commission is not a paragraph in a prompt. When an agent or its tool reaches for a domain off its list, a unix socket or another local port, the kernel and the proxy refuse before a packet leaves the machine. For example, in Experiment E4 the confined browser completed three calls in 13 seconds inside its sandbox limit (3 GB in E4, tuned to 1 GB for daily use after the C1 memory benchmark) with 0.37 ms median gate overhead, and refused an off-list host before a packet left the machine. The court does not ask the agent how it went. In an early experiment a local model’s browser tool failed, and the model fetched the page another way, wrote the correct answer and once reported plain success. The answer was right and the route was forbidden, because a proxy log captures routes the agent omitted. A court that reads the log sees that. A court that reads the report does not.
So the guiding sentence is short:
A Privateer sails with a letter of marque: confined by the kernel, checked at the gate, signed on return.
And the claim beside the fleet’s coding one (“Let them code. Make them prove it. Ship your ideas.”):
Send them out. Make them prove it. Sign what counts.
Rules that follow
Some of this is built and some is planned. I mark which.
- Nothing irreversible without my signature. Reading passes within its budget. Sending, publishing, paying, deleting and restarting a service cannot be set to “allow”, and there is no “always allow” for them. Planned: I approve each such step with a signature from a Nostr signer on my phone; the key never touches the machine. Until that is built, no rule that sends, publishes or pays is switched on.
- An honest “could not” is a valid result. A site that answers “too many requests” is not a failure of the model, and sending three more models against it helps nobody, least of all the site. Both the check and the rule that stops the chase are built, and they apply only when the logs show an external fault on the configured source with no successful required tool call on record.
- Local by default. Daily routines run on the model in my own machine. A paid model is the exception I name per job. Confidential documents run with no network at all, which the request record can show as zero lines. The routing exists. Scheduled routines do not yet.
- The verdict comes from fixed rules. For the first job, a wind and snow report, the limits are numbers I set. The model collects values. It never decides whether a day is good.
Where this sits next to the agents you can buy
I have not used either vendor product. Their columns come from their official documentation and pricing pages read on 2026-11-16, and plan details change fast. The fleet column is the state of today, gaps included.
| Grok Bot | Claude Cowork | Sovereign Fleet with Privateer | |
|---|---|---|---|
| Where the work runs | its own cloud computer at the vendor, on while you are away | Claude Desktop on folders you pick (plus cloud scheduled tasks on Anthropic servers) | your own machine, one kernel sandbox per job |
| Model | the vendor’s | the vendor’s | local first, a paid model only where you allow it |
| Your apps and logins | signs in to your tools on the cloud computer (with Vault and Secure Forms) | connectors to external apps and tools | no account connectors yet; confined browser and fetch capabilities for reading |
| Recurring work | routines on a schedule or an event | scheduled tasks | not built |
| What stops a wrong action | Auto Review permission modes, including Always allow | permission modes (Manual, Auto, or Skip) | rules outside the model: kernel sandbox, default deny per tool, call budget, domain list, non-empty effects, and hash-bound draft capture (phone Nostr signature planned) |
| How you know the result is right | the bot reports | the assistant reports | a gate decides from records the agent cannot write (hash-chained call logs, response bodies, egress logs, and schema checks) |
| Where your data goes | to the vendor | to the vendor’s model | nowhere, as long as the job stays on the local model |
| Cost | part of The X Subscription (Premium from 30 USD/month) | paid plans (Pro from 20 USD/month, 17 USD/month billed annually) | hardware and power |
| Using it | mobile and web app | desktop, web, and mobile app | chat message, issue or MCP call, CLI (fleet draft) and dashboard |
| Source | closed | closed | open |
| Strongest point | works across apps today, without your computer | lowest entry, strong model, many connectors | limits and proof that do not depend on the model, for any agent |
| Weakest point | your sessions live on someone else’s computer | the same model does the work and reports on it | does far less today, and you run it yourself |
Privateer does not replace either of them. An assistant can sit in front, talk, remember and hand over the job. What the fleet adds is the commission and the court, on your own machine, for whichever agent does the work.
What exists today
The foundation exists and is tested as of November 2026 (1354 passing tests): third-party tool servers (browser under a 1 GB memory limit and a confined fetch capability under 512 MB with a shared provenance cache) inside kernel sandboxes, one gate in front of them, two-phase draft capture capped at 5 drafts and 24 hours (with phone Nostr signature planned), hash-chained call and response body logs the agent cannot touch, and a native evidence check that verifies results and recognises an honest “could not”. How it is built, what broke on the way and what is still open is in the engineering article, and the feature status overview is on the Sovereign Fleet page.
What does not exist yet: scheduled routines, service logins, and a full run of real daily days on a live job. The first one will be the wind and snow report for my home region. When it has run for a few real days, with a record of forecast against reality, the numbers go here.