Agents
Practical agents with real execution
What happens in a run
- Step 1
Queued
Your run is stored with its input. Only one run per agent is active at a time.
- Step 2
Isolated
A worker starts a separate process with no database access, no provider keys, a memory cap, and a time limit.
- Step 3
Steps
The model plans, calls permitted tools, and reads results. Each model call is metered like an API request.
- Step 4
Result
The final answer, every step, and the total cost are saved in run history.
Tools an agent can use
Tools are opt-in per agent. There is deliberately no shell, code execution, file system, or wallet signing.
Web fetch
Read a public web page over HTTP(S). Private networks are blocked; an admin allowlist may apply.
Calculator
Evaluate arithmetic expressions precisely.
Current time
Read the current date and time.
Agent memory
Remember small notes between runs (16 KB per agent).
Asset prices
Read verified onchain prices for Stovra's allowlisted assets.
Account balance
Read your Stovra balance (read only).
Start from a template
Templates are starting points you can edit before saving.
Research digest
Reads the pages you give it and writes a sourced summary.
Tools: web_fetch, calculator, current_time
Asset watch
Checks allowlisted asset prices on a schedule and notes changes.
Tools: asset_price, memory, calculator, current_time
Spend guard
Reviews your Stovra balance and estimates runway.
Tools: account_balance, calculator, memory, current_time
Limits that protect your balance
- Each agent has a per-run spend cap. The run stops when the next model call could exceed it.
- Runs time out and are terminated by the worker, not left running in the background.
- Scheduled agents run no more often than every 15 minutes and pause automatically if your balance runs out.
- Web reads are limited to public addresses. Private networks, cloud metadata endpoints, and non-standard ports are blocked.