Runs right away
List, get, search, report, render and check. Nothing changes in your platform, so nothing stops it.
list_campaigns get_agent run_report
dylr.dev Safety
Safety tiers, arming and approvals. Reads run. Writes ask. Risky actions wait for you.
Every DylrDev tool has a tier, and our server enforces it on every call. The tier decides whether a call runs, asks you inside your AI client, or waits until you say yes in the web app, where a model cannot click for you.
Three tiers. Every tool on every platform has exactly one.
Asks you in your AI
Reversible creates and edits. The tool is labeled so your AI client asks you to confirm, and it shows the exact change first.
update_skill create_disposition
Waits for you in the web app
Deletes, starting or stopping dialing, bulk lead changes, users and roles, publishing live flows, spending money and raw API calls that are not reads. Refused until you approve that exact call or arm the connection.
stop_campaign publish_flow upload_call_list
Unknown tools are treated as dangerous, and a connector that leaves any tool unclassified is refused by the server. Tool counts by tier for each platform are on the platforms page.
Why not just have the AI ask? Because the AI is the thing that might be tricked.
A web page, an email or a ticket your AI reads can carry hidden instructions. That prompt injection can tell the model to say "the user already confirmed". Many clients also have an always allow setting, and some run unattended. A confirmation that only happens inside the chat can be skipped, forged or clicked through by habit.
So for the actions that matter most, the yes happens somewhere the AI cannot reach: you, signed in to the DylrDev web app, with a fresh password check. The server refuses the call until that yes exists.
Arming. A time window you open on purpose.
Open the connection and click Arm
Pick 15, 30 or 60 minutes. Owners and admins can also pick 4 hours. You confirm your password if you have not in the last 15 minutes.
Dangerous tools work until the timer ends
Expiry is checked on every call, not once. Every dangerous call made while armed is still logged with who armed it.
Disarm is one click
Revoking the connection also ends arming at once. With auto disarm on, an arming that sees no dangerous call for 20 minutes ends by itself.
Arm only while you are watching
If you arm and walk away, your AI can use that window. Disarm when you are done.
Approvals. Say yes to one exact call.
Nothing fails silently
When the connection is not armed, the dangerous call is stored as a pending approval and your AI gives you one link straight to it.
Single use, exact arguments
Approve just this call, or approve and arm. An approval matches only the arguments you saw, works once and expires after 15 minutes.
Batches and a second approver
Several waiting items can be approved together. Teams can require a different admin or owner than the person whose AI asked.
Safety presets. Owners pick one for the whole organization.
| Preset | What changes | Good for |
|---|---|---|
| Strict | Balanced, plus a required second approver, members cannot arm, and arming lasts 15 minutes at most. | Production tenants with several admins. |
| Balanced default | The three tiers as described above. Arming defaults to 30 minutes. Members cannot arm. | Most teams. |
| Hands-on | Owners may trust specific dangerous tools or actions so they run like writes. Each trust is shown with who set it, audited, and undone in one click. | A sandbox, or a busy admin who builds all day. |
The never-trust list
No preset can lower these: deletes, placing calls or sending messages, buying numbers or anything else that spends money, raw API calls that are not reads, changes to credentials, users or roles, and bulk changes above 25 items.
Roles. Give people the lowest role that works.
| Role | Can | Cannot |
|---|---|---|
| Owner | Everything, including presets, trusted tools and deleting the organization. | See credentials. Nobody can. |
| Admin | Manage connections, arm, approve, create tokens, invite and remove people other than owners. | Choose the safety preset or trusted tools, or delete the organization. |
| Member | Connect their own AI to the connections they were given and see their own activity. | Arm or approve, unless the organization turns on members can arm. |
| Viewer | Read the dashboard. | Create tokens or use an AI. |
The full roles matrix, generated from the live permission rules