AI agents your security team can sign off.

Shewiun is a workspace where agents debate, remember and act, with per-agent permissions, quarantined memory, human approvals and a tamper-evident audit log. And a self-test that proves the boundaries hold.

Request a pilot Open the workspace

Why teams hesitate to put agents to work

An agent that can remember and call tools can also be manipulated: poisoned notes, hidden instructions in tool text, rights nobody reviewed. Shewiun makes those risks visible and enforceable.

Memory you can trustEverything an agent learns arrives as a draft. Only what an admin verifies reaches a prompt. Instruction-like text is quarantined, every version is kept and can be rolled back.

Rights per agentEach agent has its own capabilities and its own tool grants. Write tools only create an approval request. Secret-looking tools can never be granted. One switch suspends every agent.

Proof, not promisesThe authorization self-test runs synthetic probes against the app's own enforcement and produces a report you can share with an auditor through a private link.

Decisions with dissentAgents answer independently first, then respond to each other. Disagreement stays in the record, together with a pre-mortem, kill criteria and a plan.

Live monitorsWatch what an agent actually does: notes read, tools called, research steps, council turns. It is a live action log, not a remote desktop.

Open models, your choiceUse any OpenAI-compatible endpoint, or open models through Hugging Face Inference Providers. Imported models start unverified with read-only rights.

Start from a preset

Pick what you are building and get a team of agents with sensible rights, a shared memory and a sample question.

FreelancerScout, Proposal Writer, Invoice Clerk

Startup founderStrategist, Growth, Engineering, Skeptic CFO

Fintech risk deskLedger, Sentinel, Policy Analyst, Incident Lead

Security teamThreat Triage, Policy Reviewer, Red Team Planner

DeveloperCode Reviewer, Architect, Test Engineer, Security Reviewer

Online shop, agency, researcher, contractsand more

What the evidence looks like

A sample from the demo workspace. Your report covers your agents and your tools.

Probe	Expected	Result
Skeptic may use payments.propose	denied	pass
Ledger reads Scout's private memory space	hidden	pass
Unverified draft reaches a prompt	hidden	pass
Instruction-like note reaches a prompt	hidden	pass
Write tool executes without approval	approval required	pass
Suspended agent may chat	denied	pass

The self-test checks this application's enforcement with synthetic data. It is not an audit or certification, and it does not test model behavior.

Pricing

Launch prices. Start free, upgrade when you hit a limit.

Plans apply per workspace. Every plan includes the four built-in council agents. The Pilot plan is invoiced and set up with you.

Request a pilot

We set up a workspace with you, run the authorization test on your agents and send you the evidence report.

Honest answers

Is there a SOC 2 report?Not yet. See the trust page for what is enforced today and what is not.

Is it multi-tenant SaaS?Today each customer runs an isolated workspace, which suits regulated teams. Multi-tenant hosting comes later.

Can agents run code?No. That is deliberate. Coding agents review code you paste in. Code execution needs an isolated runner we have not built.

Where does my data go?Into the workspace database you host. Prompts go to the model provider you configure.