AgentParley
LOCKED HOLD
Your keys sealed, your plugins boxed in, your agents allowed exactly what you say. Here's how.
We run your agents so you don't have to, and we built the locks in from day one: per-account encryption, a fresh sandbox for every plugin call, and agents that start with no access at all. Plain facts, no "bank-grade" fog.
THE BIG THREE
Every account gets its own data key, protected by AWS KMS. Your model keys, server passwords and plugin secrets are encrypted with it (AES-256-GCM). Once saved, a secret is write-only: the console shows 'set' or 'not set', never the value.
Every plugin call runs in its own fresh, throwaway sandbox with a memory limit and a clock. No file system, no processes, and outbound requests checked against the sites the plugin declared. You see that list before you install.
A new agent can't touch your servers, other agents or tools until you allow it, one by one, on its Access tab.
THE MANIFEST
Your account and login. Your agents' settings. Chats and their files, for your plan's retention period. Workspace files. Encrypted secrets: model keys, SSH credentials, plugin secrets. Usage records with estimated cost.
Store your card: Stripe does, on its own hosted checkout. Resell or mark up your model tokens. Keep the tokens Google or GitHub hand us at sign-in. Receive your live voice audio when you use Deepgram listening: it goes straight from your browser to Deepgram.
We don't train models on your data.
THE BRIG
Every call boxed in, and every declared address shown before you install.
Every call spins up a fresh V8 isolate and destroys it the moment it returns. Hard ceilings: 128 MB of memory, 10 seconds of CPU, 30 seconds of wall-clock time.
No file system, no processes, no loading other code.
A plugin can ask your agent's model to think for it, through our API. Your API key stays envelope-encrypted under your account's KMS key and never enters the sandbox.
Plugins make HTTP requests through our egress gateway: every request is checked against the plugin's declared host allowlist, rate-limited to 60 a minute, with a 10-second timeout and a 5 MB response cap.
We work out what a plugin does by scanning its code, so it can't claim less than it uses. You see its permissions and hosts before you install.
If a new version asks for more access than you approved, it waits until you approve it.
A plugin reads only the settings and secrets bound to it. The platform token it acts with is scoped to one account and one chat, and never enters the sandbox.
A plugin that reshapes your agents' instructions, or the commands they send to your servers, says so on its page before you install.
SHORE PARTIES
When you verify a server, we record its key fingerprint. If it ever changes, you see old and new side by side and must accept the change yourself. Direct connections to private or loopback addresses are refused.
A small program on your machine dials out to us, so you never open a port. You approve the machine by entering a code in the console.
Each agent has its own list of servers, colleagues and tools. Specialists can be locked to their own crew. Agents only ever talk to agents in your account.
Our web fetch and the chat connections we hold for plugins refuse private, loopback, link-local and cloud-metadata addresses.
THE HULL
Every record carries your account, taken from your signed login, never from what a request claims. Every record is scoped to your account in the application layer.
Customer, staff, plugin and internal-service APIs are kept apart, each with its own token audience. Automated architecture tests fail the build if code crosses between them. Staff accounts are not customer accounts.
Passwords hashed with PBKDF2 (100,000 rounds). Short-lived signed access tokens (1 hour) with rotating refresh tokens and replay detection. Sign in with Google or GitHub using PKCE.
Files are stored in encrypted S3 buckets with public access blocked. All traffic runs over TLS.
Servers and database in the United States (Hetzner, Ashburn, Virginia). Encrypted files and keys in AWS Northern Virginia. Requests to the model provider you connect go to that provider, wherever it runs. Payments go through Stripe; sign-in through Google or GitHub if you choose them.
Nightly database backups plus continuous change logs, kept for 30 days.
WEIGHING ANCHOR
Delete your account from Account settings any time. It locks immediately, you have a 30-day window to restore it, and then everything is purged. Before you go, export any chat as a transcript, download your workspace files, and export your usage as CSV.
Found a security issue? Email security@agentparley.ai.
Sign up in your browser, paste an API key from the AI provider you choose, and your first agent is waiting. Add a crew, a plugin or a face whenever you like. We run the servers and ship the updates.
