How we use AI

How Missived is built

Missived is designed and directed by its founder, Cody West. Much of its code is written with AI coding assistants — Claude Code by Anthropic and OpenAI Codex — working under his direction. Cody decides what to build and is responsible for what ships.

Every change goes through automated tests, continuous integration (CI) checks and review before it ships. Those checks help catch mistakes; they do not make software infallible. The self-hosted edition's code is public under the AGPL-3.0 licence so anyone can check the work. See self-hosting and what we have tested.

No AI inside your workspace unless you connect one

Missived has no built-in AI reading, summarising or ranking your messages or files. We do not train models on your content. No AI features run on our servers, and our servers never call an AI service.

AI touches your workspace only if you choose to connect an assistant through MCP (Model Context Protocol), or install an app that uses AI. That gives the provider you chose access to the content its permissions allow. Its own privacy terms apply to what you send it.

An assistant is limited by your own access, the scopes you approve and your workspace's settings. You can choose read-only access or allow actions. Inviting people and sharing outside the workspace, and changing settings and roles, need separate permissions. With those permissions, an assistant can act on your behalf; check your Activity for what it did. A workspace Owner can require confirmation before an assistant shares outside the workspace.

Signed-in assistants cannot read direct messages. A personal token can read your own DMs only if you explicitly tick “Include my direct messages”; app tokens cannot. Workspace Owners and Admins can switch MCP access off. Disconnect an assistant in Settings → Connected apps, or revoke its token in Settings → Developer.

Read the MCP guide, its permission scopes and the developer docs before connecting one.

Bug reports: only the assisted lane you choose

When assisted reporting is enabled, the bug-report form says an AI assistant helps us fix bugs and asks you not to include sensitive information. Choose Send to our team only to keep your report out of the assistant lane. When the pipeline is off, all reports are team-only.

The founder's own Claude routine fetches only reports in the assisted lane through a narrow connection. Our servers never send reports to an AI service themselves. Before the routine can read a report, we strip identifying details again, including names, emails, links, secrets and workspace names. The routine has no access to your account, messages, files or calls through that connection. Message content is not collected automatically; please keep it out of descriptions and screenshots too. Stripping cannot guarantee that every sensitive detail in free text is caught.

You choose whether to attach a screenshot and can crop or blur it in your browser before sending it. The edited image is encrypted in storage; the routine can fetch it separately when needed. Screenshots are deleted 30 days after the report is resolved, and reports and screenshots are removed with your account. Report text otherwise remains in our database today; a shorter retention period is still to be decided.

GitHub receives a neutral summary of the problem, never the raw report, your words or a screenshot. The code-fixing routine works from that summary and the repository. We use Claude under terms or settings that do not allow training on these reports. Anthropic is listed on our privacy promise.

Server upgrades: a procedure, with approval

Claude helped design and test the upgrade procedure to reduce downtime and make the steps consistent. The Console upgrade flow is designed to run inside Missived: it shows the proposed server, price and expected downtime, asks for the founder's approval, then performs the upgrade and checks the service afterwards. Claude is optional, not required to approve or run that flow.

This procedure uses server capacity readings, infrastructure quotes and upgrade status, not customer messages or files. Helping build the procedure does not give an assistant access to your workspace.

When we diagnose a problem, an AI assistant may look at server logs and error details — only after IP addresses and identifiers are removed, and never message content.

Feature ideas: coming soon

We plan an optional way to suggest an idea and have a Claude routine group similar suggestions for Cody to review. It is not available yet. The plan follows the bug-report rules: a clear disclosure and choice on the form, identifying details stripped before assistant review, no automatic collection of message content, and the same screenshot controls and deletion rules. We will explain the flow before enabling it.

Who sees what

Who	What they can see
Coding assistants helping build Missived	The code and development material used for the task. Building the product gives them no workspace access.
An assistant or app you connect	Content and actions allowed by the permissions you grant and your own access. You choose the provider and can disconnect it.
The founder's Claude bug-triage routine	Only assisted reports, stripped again before review, and edited screenshots when needed. Never team-only reports or automatic access to conversations.
GitHub and the code-fixing routine	A neutral bug summary and the repository, not raw reports or screenshots.
AI assistants helping diagnose problems	Server logs and error details only after IP addresses and identifiers are removed, and never message content.
Missived's upgrade flow	Server capacity, infrastructure quotes and upgrade status. No customer messages or files; Claude is optional.
A future feature-idea routine	Not active yet. The planned rules match assisted bug reporting.

Questions? Email [email protected]. Our privacy promise names the providers involved and explains retention, access and your choices.