Security

Built to be trusted with your team's work

Collabicle holds your docs, projects, data, and conversations — and opens them to AI agents. That only works if the boundaries are real. Here's how the system is built, in plain language.

Workspace isolation

Workspaces are hard boundaries — separate members, data, and billing. Every read and write is checked against workspace membership on the server, on every request; there is no client-side trust.

Authentication

Password sign-in with hashed credentials and OTP-verified resets, or Google OAuth — we never see your Google password. Sessions are managed server-side and revoked on sign-out.

Role-based access, everywhere

Admins and members have distinct capabilities, enforced in the backend functions themselves — not in UI conditionals. The same role model gates humans, agents, and automations alike.

Sandboxed automations

User automation code runs in an isolated sandbox, not on shared app servers. Workspace secrets are injected at runtime, scoped to the workspace, and never exposed to other tenants.

Payments by Stripe

All payment details are collected and stored by Stripe, a PCI-DSS Level 1 provider. Card numbers never touch Collabicle's servers — we hold only your plan and billing status.

Encrypted infrastructure

All traffic is TLS-encrypted in transit and data is encrypted at rest by our infrastructure providers. Collabicle runs on SOC 2-audited platforms (Convex, Vercel).

The part most tools get wrong

AI agents without god mode

Giving agents workspace access is a trust decision. The MCP server is architected so you can say yes carefully — and prove what happened afterwards.

OAuth 2.1, not API keys

Agents authenticate through the same auth system as people, tied to a real workspace identity you can see and revoke instantly. No long-lived keys floating around in config files.

Deny-by-default tool gating

The MCP tool list itself is filtered by the caller's role before the agent ever sees it. A tool an agent isn't allowed to use doesn't error — it doesn't exist for that agent.

Identity that can't be forged

The acting identity is injected server-side by the gateway and stripped from client input. No tool call can claim to be a user it isn't — by construction, not by convention.

A full audit trail

Every agent tool call is logged: which identity, which tool, which workspace, when. Usage is visible per workspace, so agent activity is never a black box.

What this looks like in practice

$ connect collabicle —mcp
✓ OAuth 2.1 · role: member
tools/list → filtered by role
✗ admin tools: not present
AgentActionWhen
Claudeboard_write2m
Claudewiki_read9m
Research bottables_read12m

Role-filtered tools on connection; every call in the audit log after. Read the full architecture on the MCP page.

Data practices, in writing

What we collect, who processes it, how long we keep it, and your rights over it — documented plainly. We don't sell your data, and we don't train AI models on your content.

Found something? Tell us.

If you believe you've found a vulnerability, we want to hear about it — before anyone else does. Email us with steps to reproduce; we read security reports first and respond fast. Good-faith research is welcome; we won't pursue action against it.

ibrahim@collabicle.com

An honest note on maturity: Collabicle is a young product built on audited infrastructure. We don't yet hold our own SOC 2 certification — when we do, it will be announced here. If your team has specific security requirements, email us and we'll answer directly.

Boundaries you can verify, not just believe

Start free, connect an agent, and watch the audit log record every move it makes.