Enterprise edition · Preview
Cohandry for organizations on AWS
The Enterprise edition is the same desktop app your developers already use, except that it takes its AI access, its safety rules and its audit trail from your company's AWS account instead of from personal API keys. People sign in with their company identity, every model call goes through Amazon Bedrock under rules your administrators set, and your security team can find each call in AWS CloudTrail under the name of the person who made it.
The Enterprise edition is in preview and is not part of the public download yet. We are testing it end to end on AWS before release, so if your organization would like to evaluate it, email us and tell us how your teams use AWS today.
What it adds to the free app
Company sign-in instead of API keys
Developers sign in through AWS IAM Identity Center, which is the single sign-on many companies already connect to Okta or Microsoft Entra ID. Cohandry then receives short-lived credentials for the role your administrator assigned, keeps them only in the app's local coordinator, and never passes them to the agents doing the work. When a session ends or someone leaves the company, their model access stops, while their local projects stay on their computer.
Models from your Amazon Bedrock account
Agents call models through Amazon Bedrock in the region your administrators choose, including OpenAI's open-weight gpt‑oss models, and the usage appears on your AWS bill. Because your policy lists the approved models, an agent cannot quietly move to a model that your company has not reviewed.
AWS-enforced guardrails
Every model call carries your company's Amazon Bedrock Guardrail, and the IAM permissions behind each person's sign-in deny any call that does not, so the requirement is enforced by AWS rather than by a setting in the app. Cohandry also checks file contents and tool results before a model sees them: text that looks like a prompt injection is withheld, and email addresses and phone numbers are masked.
Guardrails reduce these risks rather than remove them, which is why Cohandry's own permission checks, command approvals and reviews still decide what agents are allowed to do.
Company knowledge, when a project needs it
Directors and reviewers can search your company's documents through an Amazon Bedrock Knowledge Base, and each passage is checked by the guardrail before it reaches a model. Knowledge search is switched on per project, because it relies on documents your company stores in its own AWS account.
IT-controlled policies
Administrators install one policy file that ordinary users cannot edit. It sets the region, the approved models, the guardrail and the knowledge base, allows only the company's Bedrock connection, and stops tasks that would run on personal accounts. It also configures your proxy server and corporate root certificates. If the file is damaged or can be edited by users, Cohandry stops making model calls instead of ignoring it.
CloudTrail audit trails
Each model call is recorded in AWS CloudTrail under the signed-in person's name, together with the guardrail version it carried. Your security team can therefore review AI usage with the same tools and retention rules it applies to everything else in AWS.
Security and data handling
Security teams usually ask the same questions about an AI tool: where the data goes, who controls access, what an agent is able to do, and what evidence exists afterwards. These are the answers for the Enterprise edition as it stands in the preview.
Your data stays on your computers and in your AWS account
Projects, source code, files, Git repositories, records and task history stay on each person's computer, in the project folder and the app's local database. The app sends none of it to Cohandry: it contacts cohandry.com only to check for updates, and to verify Packs if someone connects a Cohandry account, and it has no telemetry. What leaves the computer is each model request, which contains the task and the files or tool results an agent needs for that step, and it goes only to Amazon Bedrock in your AWS account and the region your policy names. AWS states that model providers, OpenAI included, have no access to Bedrock prompts and completions. Company documents for knowledge search live in an encrypted S3 bucket in your account, and only projects that switch knowledge search on can query them.
IAM governance with least privilege
Each person's access comes from a permission set that allows only the approved models in your region, your guardrail and your knowledge base. It explicitly denies any model call that does not carry the guardrail, and it cannot change or delete the guardrail itself. Your organization can add service control policies on top, your administrators set how long a session lasts, and multi-factor authentication follows your identity provider.
No standing credentials on laptops
The app holds only short-lived role credentials, keeps them in the memory of its local coordinator, and never writes them to disk or passes them to agents, tools or other computers. When a session ends or a person signs out, model access stops at once, and work in progress pauses until they sign in again rather than failing or retrying.
Controls inside the app
Agents act only through the coordinator, which checks every requested action against its permissions. File tools are limited to the project folder, commands wait for an approval, and nothing is accepted until reviewers have seen it and the recorded checks have passed. Pushing code, opening pull requests and merging remain human actions, and the desktop app runs with Electron's context isolation and sandboxing.
Evidence for your security review
CloudTrail records each model call under the person's name with the guardrail version it carried, and each guardrail intervention also appears in the project's activity history. Your team can inspect the IAM policy directly and test it with the IAM policy simulator. Knowledge base searches and standalone guardrail checks are logged as well once your trail records Bedrock data events.
What we have not done yet
Cohandry has not completed an independent security certification such as SOC 2, it has not been tested in AWS GovCloud (US), and today's builds are not yet signed with an Apple Developer ID or a Windows code-signing certificate. We will say so plainly in every evaluation until that changes.
How an organization sets it up
Deploy the AWS stack
An administrator deploys one CloudFormation template. It creates the Cohandry group and its least-privilege role in IAM Identity Center, the guardrail together with the IAM rule that requires it, and an encrypted S3 bucket with a Knowledge Base for company documents.
Install the policy file
A short script turns the stack's outputs into the organization policy file, which IT then installs through its device management tools in a location ordinary users cannot change.
Add people to the Cohandry group
Access follows group membership in IAM Identity Center, so adding or removing a person there is how a seat is granted or withdrawn.
Sign in from the app
Each person installs Cohandry, chooses Sign in with your organization, and approves the request in the browser with their usual company login.
What we are building next
Before the Enterprise edition is released, we plan to add seat management for administrators inside the app, an export of agent actions and human approvals to your own S3 bucket or CloudWatch, macOS builds signed with an Apple Developer ID and notarized, a signed Windows installer, and support for existing AWS profiles and Bedrock inference profiles.