Security and limits
Where an agent’s keys live, how the AI Gateway keeps provider keys off the agent, what a public agent cannot do, and the limits that apply.
Provider keys stay in the AI Gateway
No model provider key is ever stored on an agent.
- By default each agent gets its own AI Gateway, a service in the same environment. Your provider key is kept in the gateway's variables.
- The agent gets its own gateway key, limited to its one model. That key is its only model credential.
- StackBlaze reaches the gateway's admin API privately, never over a public URL. The gateway's master key is never returned, logged or written to the agent.
- An agent on an AI model of the same environment connects to it like any app.
Inside the agent, each chat process gets only its model key and its own settings, not the rest of the service's environment.
Agent keys and secrets
| Secret | Prefix | Where it is kept |
|---|---|---|
| API key | sbag_ | Shown once. The agent keeps only a SHA-256 hash. |
| Webhook secret | sbwh_ | Shown once. The agent keeps only a SHA-256 hash and compares it in constant time. |
| Integration token | (the service’s own) | In the agent’s variables, masked. Never returned. Sent only to its own server. |
Revoke a key or webhook from Triggers when it is no longer needed. Message text is never logged.
Open chat signs you in with a link that is valid for 120 seconds. Each team member gets their own session.
Private and public agents
An agent is private by default. Only your team (with Open chat), callers with an API key or webhook secret, and its schedules start chats.
Turning on the Website widget makes it public: anonymous visitors can chat with it. A public agent:
- has no shell or file-editing tools:
bash,editandwriteare refused, whatever the settings say; - has no integrations, because a visitor could talk the agent into using them;
- is shown only inside the sites you list, and visitors can only continue their own chats.
The widget is not the service's network access. It does not change who can reach the agent's URL; API keys and webhook secrets still protect the API.
What the tools can reach
File tools only reach the agent's knowledge folder (and, for agents with write tools, the chat's scratch folder). URLs, internal addresses, ~, system folders and links that point outside are refused. Skills can be read, never written.
The tools you can turn on are read, grep, glob, web_search, todo, bash, edit and write. Tools that run code or reach outside the knowledge folder in other ways are not offered.
Who can change an agent
Agents follow the project's team permissions. Reading an agent, its sessions and its files needs read access. Creating, changing, deleting, starting runs and Open chat need write access.
Limits
| Limit | |
|---|---|
| Instructions | 20,000 characters |
| Message | 8,000 characters |
| Turns per chat | 30 |
| Reply time | 120 seconds |
| Live chats at once | 5 |
| Chat history | Removed after 90 days without an update |
| Knowledge | 10 MB per file, 200 files, 200 MB |
| Skills | 50; steps up to 40,000 characters |
| Integrations | 10 |
| API keys | 20 |
| Webhooks | 20; 64 KB body; 60 calls per minute each |
| Schedules | 20; message up to 4,000 characters |
| Widget | 20 sites; 10 messages per visitor per minute; 60 messages and 10 new visitors per IP per minute; 3 new chats per visitor per day |
| Evals | 20 per agent; 100 test cases, 20 checks, 3 epochs, 300 trials, 600 s per trial; last 20 runs kept |
| Optimizations | Budget $1 to $500; one run at a time; last 10 runs kept |
The agent and its gateway are services, so your plan limits apply to them too.