DataShield Commander v3.0.15
Commander: the operations cockpit for coding agents
One tab. Every agent. Commander is a self-hosted Claude Code cockpit: the one place your Claude Code sessions live. You launch each one from a persona, with a boot prompt you approved. Commander keeps its transcript. A dev, test, prod ladder promotes finished work. That ladder stops at a dirty tree. It asks a human for a reason.
It runs on your box, against your PostgreSQL. Your transcripts are not a feature of somebody else's cloud.
Token Analytics
Every token, by model, by project, by account. Before the invoice tells you.
Model spend is the fastest-growing line in most engineering budgets. Run ten agents in parallel, and tokens burn ten times faster than one. Half of the teams shipping AI products still do not track it at all. The bill arrives. The CFO asks which project did that. The honest answer is a shrug.
Commander meters every session it runs. Token Analytics shows tokens in and out, cache reads and writes, turns, and dollar cost per model. It filters by project and by account. The hourly view covers the last 24 hours. It hatches each model's sub-agent share too. So a fleet that fans out at 03:00 shows up as exactly that. It works the same in the Free edition as in Team, on your own machine. The transcripts and the meter live in your own PostgreSQL. No enterprise analytics tier required.
Editions
Same Commander in every box. What changes is how many humans and agents it serves, and where it runs. What also changes is whether you may charge other people for what you build on it.
Free
One person, one laptop, finding out what a cockpit is for.
Docker image · your machine
- 1 project, 2 concurrent sessions
- Built-in personas and prompt library
- Bring your own Claude subscription or API key
- Transcripts kept in your own PostgreSQL
- Community support
- Personal, non-commercial use
Personal
One engineer, real projects, more agents than hands.
Docker · or $100 / month on a DataShield-provisioned private server
- Unlimited projects and sessions
- Custom personas and boot prompts
- Promotion ladder with audited override
- Screenshots, artifacts and git tracking
- Attach to daemon sessions, transcript search
- 1 user · internal use
Team
Five people who would like to stop asking "which session did that?" in Slack.
Docker · or $500 / month on a DataShield-provisioned private server
- Everything in Personal, for 5 users
- Sub-agent fleet with nesting blocked
- Several Anthropic accounts, chosen per persona
- Encrypted integrations vault
- MCP servers brokered through Auth
- Deployment history and audit trail
- More seats on request
Enterprise
Companies running agents in production, and anyone selling what they build.
Private air-gapped cloud, your hosting, or ours
- Everything in Team, for your whole organisation
- Commercial use and white-label rights
- SSO with your identity provider
- Signed license plane and fleet state
- Premium features and priority support
- Marketplace listing for your product
Selling anything built on Commander means Enterprise. Agency work included.
Prices are per month, billed monthly, plus applicable tax. Docker editions run on hardware you provide. DataShield-provisioned means a dedicated single-tenant server we operate for you, sized for the edition. Your model provider bills model usage, not us. We are not in the token resale business.
What each edition includes
| Free | Personal | Team | Enterprise | |
|---|---|---|---|---|
| Users | 1 | 1 | 5, more on request | Organisation |
| Projects · concurrent sessions | 1 · 2 | Unlimited | Unlimited | Unlimited |
| Personas and boot prompts | Built-in | Custom | Custom + sub-agent fleet | Custom + sub-agent fleet |
| Promotion ladder | no | ◆ | ◆ with history and audit | ◆ with history and audit |
| Anthropic accounts | 1 | 1 | Several, per persona | Several, per persona |
| Integrations vault · MCP via Auth | no | no | ◆ | ◆ |
| SSO (OIDC) | no | no | no | ◆ |
| Hosting | Docker | Docker or provisioned | Docker or provisioned | Air-gapped, yours, or ours |
| Commercial use and white-label | no | no | no | ◆ |
| Support | Community | Email, next business day | Priority, named contact |
Agent fleets sharing the same understanding
One Claude Code session in a terminal is fine. You can see it. You can Ctrl-C it. Ten of them across four repos is a different sport. Terminals hide behind other terminals. Someone asks which session ran the migration. The honest answer is a shrug. The boot prompt explained how this project works. It lived in a shell history someone cleared on Tuesday.
Commander is the room where all of that runs. Every session is a tab with a name. Each launch starts from a persona you wrote once. Every transcript lands in a database you own. An agent claims the deploy is done. The promotion ladder tells you if that's true.
Three jobs Commander does every day
Run the fleet
Many Claude Code sessions run per project. Each is a real terminal in the browser. Not a chat widget doing an impression of one. Resume a session. Attach to one still running in the Claude Code daemon. Pick up mid-thought. Search the logged transcript from the top bar. Model, effort level and account sit in the breadcrumb. Nobody has to ask whose API key is on fire.
Launch with a persona
The Agent Launcher starts a session from a persona. Pick Developer, Tester, Release Manager, DevOps Engineer, Deep Research. Or the one you wrote at 2 a.m. and are quietly proud of. A persona carries a boot prompt, a model and effort setting, skill packs and MCP bindings. Personas double as a sub-agent roster, with nesting blocked. An orchestrator cannot spawn orchestrators. We checked.
Promote with evidence
The Deploy tab is a per-project ladder: dev, test, prod, or whatever your repo actually has. Promotion fast-forwards one rung at a time. It runs migrations fail-closed, restarts, and health-checks the target. A dirty tree halts the ladder. An operator can override it once. That takes a written reason. It lands in deployment history, where future you will read it.
How a change earns its way to prod
The ladder is server-side and per project. Nothing an agent says in a transcript moves code up a rung. Only a promotion that passed every step does, but confident the summary sounded.
A dirty file that the promotion itself would change can never be overridden. Everything else an operator can override, once, with a reason that stays in the record.
What it looks like
Captured from a running Commander instance. Dark theme, no mockups, no stock photos of people pointing at a whiteboard.
Personas, not prompts in a text file
Pick a persona, see exactly what it boots with, launch. You see model, effort, boot prompt, skill packs, MCP servers and the sub-agent gate before it spends the first token.
Every session, kept
Logged Claude sessions per project, with prompt counts, transcript size and the persona that ran them. Reopen the conversation or read the transcript without poking the live terminal.
The tree the agent is editing
Working-tree status and a file browser sit next to the terminal. You can see what changed before you believe the summary. Commit history and a branch list with worktree collision flags live in the same panel.
Signed plan, narrowing toggles
Features light up from a signed license. The toggles here can only narrow what the plan grants, never widen it. Same entitlement vocabulary on every surface, so there is no secret door.
Claude Code commands, one click away
Interrupt, compact, check plan usage, run a code review, switch model. Each action drops the command into the agent's prompt, so nobody has to remember whether it was /cost or /usage.
Docs that ship with the build
Versioned in-app docs cover sessions, agents, the ladder and Claude Code auth. Our favourite line is the isolation note: your data lives in your database, on your server. Full stop.
Underneath the self-hosted Claude Code cockpit
The parts you only notice when they are missing.
Project and work tracking
Each project points at a git tracking repo: bugs, features, releases and session updates as plain markdown. Markdown does not need a login. A boot-sequence viewer hands agents the same context a new engineer would read on day one. Commit detail opens any change in one click.
Accounts and credentials
Several Anthropic accounts side by side, OAuth or API key, chosen per persona and shown on every session. Third-party credentials sit in an encrypted store, masked by default. A health probe finds the dead token before an agent does.
Screenshots and artifacts
Paste a screenshot into a session. Commander files it against the project. An annotation tool bakes your circles and arrows into the PNG. Commander also captures anything an agent produces from the transcript. It goes into a provenance registry and opens in a sandboxed viewer.
Process monitoring
A server tile shows real CPU utilization and hypervisor steal. It also shows status for every process the stack runs, PM2 or systemd. Start, stop and restart go through a validated command list. There is no free-form shell hiding behind a button.
MCP, brokered not held
Personas bind to the MCP servers they hold entitlements for. Commander injects that binding at launch. The catalog, credentials and entitlements stay with Auth. Commander never holds a raw MCP key. That's the kind of sentence you want to say in an audit.
Configuration, not code
Every tunable travels one chain: environment variable, settings table, client config. The server holds the authority. The browser just renders what it gets, and hardcodes nothing. That is why a Claude Code version skew shows up as a plain-language toast, not a silent 3 a.m. failure.
Building something to sell on Commander? That's Enterprise -- and we can help you Go-To-Market.
Free, Personal and Team are for running your own agents on your own work. The moment somebody else pays for what Commander runs or produces, that is commercial use. A SaaS tool built on it. A hosted agent you sell seats to. An agency delivering client work through it. All three are commercial. All three run on Enterprise terms.
Enterprise terms are simple. Commercial products built on Commander ship on the DataShield plane. You get white-labelled hosting under your brand. You also get the signed license plane. It lights up features for your customers. You get process supervision and fleet state for every instance you sell. And you get Auth. Each customer's agents leave an evidence chain you can hand them. You build the product. We run the parts nobody thanks you for, at industry-standard publishing and hosting rates.
The DataShield marketplace lists products built this way. Companies and people subscribe directly. You set the price. We take the platform fee. Licensing and hosting come already solved. They share the same plane the product was built on. Cheaper to start, longer to pay off, and no second vendor to onboard when the thing works.
Where Commander sits in the DataShield stack
Commander is the operator's seat. Auth is the authority. It signs the license that lights up Commander's features. It also owns the MCP entitlements a persona can use. Ontology is the data plane those agents work against. Corpus serves the documentation, agent definitions and skills they load.
Commander also publishes a signed manifest of the agents it runs. Auth pulls it and governs it. That manifest carries its own signing key, kept deliberately separate from the key that signs deployments. An agent roster and a release are different trust decisions. We keep them apart.
Plan enforcement is shipped and switched off by default while early deployments settle. Every plan feature is visible today. The gate exists for the day you want it shut.
Commander stays a self-hosted Claude Code cockpit in every edition. Nothing about your sessions leaves the machine you run it on.
Questions operators ask first
Where do sessions and transcripts live?
In a PostgreSQL database on the server that runs Commander. Not on Anthropic's servers, not in a personal Claude web account, and not with us. The only thing that leaves the building is the API call to the model.
Is it only for Claude Code?
Claude Code is the first-class provider today, including attach-to-daemon sessions and the command menu. Other providers still work. They have not had the same love. We would rather tell you that than sell you a checkbox.
How does it stop an agent from promoting broken code?
The ladder is server-side and per project. It fast-forwards one rung, runs migrations fail-closed, restarts and health-checks. A dirty working tree halts it. If an operator overrides, the reason is mandatory. The override lands in deployment history as its own audited step. A dirty file that the promotion would also change can never be overridden.
Can I run it on my own hardware?
Yes. Commander runs self-hosted, in your own cloud or data center. Or on a dedicated single-tenant server we operate for you. It needs Node, PostgreSQL and a route to the model provider. There is no multi-tenant SaaS and no shared control plane.
What happens when Claude Code ships a breaking change?
Commander pins the behaviours it depends on. It watches session output for the known skew signatures. So a CLI that is too old for a model shows up as a plain-language toast. Same if one refuses a resume. You do not get a cryptic one-liner. When a Claude Code release flipped a trust dialog to default to exit, the fix shipped the same week. It is in the release notes, with the swearing removed.
Can agents launch their own sub-agents?
Only if you allow it per persona. Two flags decide who may be a sub-agent, and who may use them. Both are off by default. Sub-agents cannot spawn sub-agents. We deny the Agent tool one level down, without exception. We have all seen what happens otherwise.
What counts as commercial use?
Anyone paying for something Commander runs or produces. A SaaS product built on it, a hosted agent you sell access to, or an agency delivering client work through it. Agency use is commercial use. Building your own company's software with it is internal use. That is what Personal and Team are for.
Do I need to host with you?
For your own use, no. Docker editions run wherever you like. DataShield-provisioned means a dedicated single-tenant server we operate for you. For commercial products, yes. They ship on the white-labelled DataShield plane, and that is the deal.
What does the Free edition leave out?
It is the same product, capped at one project and two concurrent sessions. Built-in personas only. No promotion ladder. It is for finding out whether a cockpit changes how you work. When it does, Personal removes the caps. You can stop closing tabs to make room.
Can Claude Code run multiple sessions at once?
Yes. Commander runs many Claude Code sessions per project, each a real terminal in the browser. Every session is a tab with a name, a persona and a kept transcript. The Free edition caps you at one project and two concurrent sessions. Personal and above are unlimited.
How do I promote a Claude Code agent's work from dev to prod?
Use the Deploy tab. The ladder is server-side and per project. It fast-forwards one rung at a time. It runs migrations fail-closed, restarts, and health-checks the target. A dirty tree halts it. An operator can override once, with a written reason that lands in deployment history.
Commander v3.0.15 is in production. Documentation. It runs alongside Auth, Ontology and Corpus.
See your own agents in the cockpit. A quote takes a minute. A walkthrough takes twenty.
Get an instant quoteYou've seen the proof
Ready for a number? Scope your deployment and we'll price it against your own economics.
Get your quote →