Tool reference
The agent has a tool palette. Stratam picks which one to call based on what you're asking — you rarely have to name them explicitly. Here's the full list, what tier unlocks each, and when the model tends to reach for it.
Universal (all paid tiers)
remember
Save a fact to your vault. Stratam fires this automatically when you share something worth keeping (preferences, names, decisions, project notes).
Params: content (string, required), tags (string[], optional)
When the model picks it:
- You say "remember that …" or "save this for later"
- You share a preference or decision the agent should reuse
- You give it a name/email/role that'll come up again
recall
Search the vault for past notes. Uses hybrid search (PostgreSQL full-text + pgvector semantic similarity), so "what did I say about pricing last week" finds the right entry even without exact keywords.
Params: query (string, optional — empty = newest), limit (1-10, default 5)
current_time
Returns the current UTC time and the user's local time. Stratam calls this before answering anything time-sensitive ("when's my next meeting", "how long ago did I email Acme").
web_search
DuckDuckGo for real-time info: news, prices, current events, fact-checking. Returns ranked snippets; the agent synthesizes them into the reply.
When the model picks it:
- Question references "today / latest / current / price of X right now"
- Topic is too new for training data
- User explicitly says "search for …" or "what's the news on …"
Operator+ tools
schedule_check
Create a recurring or one-shot monitor without leaving the chat. You say "ping me every hour about BTC" and the agent fires this with a parsed schedule.
Params: label, instruction, schedule_kind (interval | cron | one_shot), schedule_spec, notify_channel, notify_target?
http_fetch
GET (or POST) a URL and return the response body. Used for scraping, hitting APIs, polling health-check endpoints.
Params: url, method? (default GET), headers?, body?, max_chars? (default 5000)
Output sanitization: Bytes returned here get scrubbed for prompt-injection patterns before re-entering the agent loop. See docs/best-practices for the threat model.
browser_visit
Headless Chromium browse — for sites that need JavaScript to render. Slower than http_fetch (5-15s per page) but works on dynamic SPAs.
Params: url, wait_for_selector?, extract_text? (boolean — strip HTML), max_chars?
code_sandbox
Run Python code in an isolated Docker container. 30-second timeout, no network access, 256MB memory cap. Used for math, data parsing, ad-hoc analytics.
Params: code (Python source), stdin?, timeout? (max 30)
Builder+ tools
delegate_to_agent
Hand off a subtask to a specialist agent from the 244-agent squad. Each specialist has its own prompt + tool palette tuned for a domain (Forge for code, Scout for research, Closer for outbound, etc.).
Params: agent_key (e.g. forge, scout, trader_x), task (string)
Browse the catalog at /agents and enable the ones you want available for delegation. The user's enabled set gets injected into Stratam's system prompt so the model knows what's on offer.
Custom agents
Builder tier can define their own agents (name + system prompt + tool checklist + optional model override) from /agents → + Custom agent. Custom agents work the same as built-in ones — they're delegate-able with delegate_to_agent using custom:<uuid> as the key.
Tool gating
Every tool runs through two checks:
- Tier gate in
tools.available_tools(user)— the user's tier dictates what's in the catalog the model sees - Cap gate in
usage.is_over_cap(user)— if you're over your monthly action cap, the assistant returns a canned response instead
If you belong to a paid org workspace, the cap is pooled across all members rather than personal. The owner gets the 80% warning email instead of individual members.
Cost per tool call
Most tools are free at the action layer — they just route data. The cost is the LLM call that wraps the tool:
- Chat replies on Sidekick/Operator use Haiku (~$0.001 per turn)
- Builder uses Sonnet for web/email replies (~$0.005 per turn)
- Code sandbox + browser visit are the longest-running tools but still under the per-turn budget
See /analytics for your real cost curve.