Safety and limits¶
Transica runs AI agents against real websites and serves tools that act for real people. These are the guardrails, all on by default.
Agents on your website¶
- Verified sites only. Agent tests and explores run only on sites you've verified, and verification is re-checked before every test.
- Read-only when signed out. With a test account, agents may only make the changes a task needs and you allow.
- Never run: payments, account closures, credential changes and messages to other people. The browser blocks them even if the agent tries.
- Contained browser. A fresh browser per run, restricted to your domain, through an egress proxy that refuses private network addresses.
- Bounded runs. Up to 20 steps and 4 minutes per run (explore agents: 12 steps and 2 minutes); stuck sessions are stopped and refunded.
- Sign-in done carefully. Redirects that resend credentials, rejections and challenges stop setup instead of retrying. Write access is allowed only once Transica has confirmed the session.
- Clean-up after write trials. Records created in comparison write trials are removed with the paired delete tool you approved and confirmed by a read. If anything can't be traced, tests that change data pause on that site.
- No bypassing. Robots rules, bot challenges and access restrictions are respected, never worked around. Public scanning isn't penetration testing.
Hosted connectors¶
- Reads go live; changes wait for approval. Tools that change data run only once an admin approves them for the served version. High-risk endpoints are never recommended, drafted into tools, or served.
- Sign-in through your provider. AI apps never see your customers' passwords or your provider's tokens; they hold only connector-bound tokens. Provider tokens are encrypted under a key only Transica's issued token unwraps.
- Rate limits. Sign-in routes are rate-limited. Signed-out calls are limited per address (300 a minute) and per connector (1,000 a minute), with a daily cap per plan. Signed-in calls have no per-address limit, because AI apps call from shared addresses. Signed-out writes you approve are capped at 1,000 calls a day per tool.
- Bounded calls. Each tool call has 15 seconds to answer.
- Turning hosting off wipes that connector's AI-app clients, grants and tokens.
Public scans and checks¶
| Limit | |
|---|---|
| Page fetched | HTML up to 750 KB, 8 s timeout plus one retry up to 45 s, up to 3 redirects (robots re-checked at each) |
| Extra pages sampled | Up to 3 |
| JavaScript and forms | Never run, never submitted |
| Free scans | 50 a day per address, 100 a day per site, 10,000 a day across everyone |
| MCP directory check | Signed out only, never calls a tool, follows no redirects; 50 a day per address, 100 a day per server |
| Connector-run scans and checks | Up to 5,000 each a day across all users |
| Addresses | Public only: private networks, localhost and sign-in-only addresses are refused |
All outbound fetches go through guards for SSRF, redirects, response size, timeouts and quotas.
Product limits¶
| Limit | |
|---|---|
| Tasks per test | 3 (task sets hold up to 10) |
| Test accounts per site | 10 |
| Endpoints per design / tools per version | 50 (advice: under 25) |
| API spec size | Up to 1,000 operations |
| Sources per build | Up to 10 previous runs |
| Explore | 20 runs, 5 minutes, site map of 30 pages |
| Comparison | 3–10 trials per side, up to 3 AI models |
| Access log upload | 50 MB, .log or .gz |