Data & AI
LLM + AI Agents
LLM-powered generation, evaluation, and research inside agent workflows — works with OpenRouter, Ollama Cloud, or any OpenAI-compatible API.
Auth & credentials
API key / token
API Token
Encrypted at rest, limited to the agents and tools you authorize, and not logged in plaintext.
Enable LLM only for the agents that need it.
Connect: Used as the Bearer token for LLM requests. Changing supported provider settings does not require reconnecting.
What it touches: Read-only model calls via your API token. No write or destructive actions.
What your agents can do with LLM
Concrete jobs an agent team handles with LLM, each running on a schedule or on demand.
Generate and revise drafts with a chosen quality level
Evaluate and classify text with confidence scoring
Extract structured data from unstructured text
Summarize long documents and synthesize researched answers
Built for agent teams
Every OrgSDK integration ships with the same guarantees — this is what makes LLM useful, not just connected.
Agents don't just answer chats. They run on a schedule — 24/7 — so LLM work happens even when you're asleep.
Agents can hold any irreversible LLM step behind a human sign-off using the built-in approval action. Add it to a workflow where it matters — approvals are an explicit capability, not enforced on every call.
Grant LLM only to the agents that need it. Credentials are encrypted at rest and surfaced only through an agent's enabled plugin functions when a workflow needs them.
Your LLM credentials are encrypted and stored per-org. OrgSDK never reads or logs them in plaintext.
Capability safety — 7 actions
Grouped from each function's source metadata. Read-only actions are non-mutating; write actions create, update, or delete data.
Read-only · non-mutating
7evalBooleanUse AI to evaluate if text matches given criteria with few-shot examples. Returns { result: boolean, confidence: number }. Requires positive and negative examples to disambiguate the classification boundary. Use confidence thresholds: >0.9 auto-process, 0.7-0.9 log for audit, <0.7 flag for human review. Optional quality parameter: 'high' (default), 'medium', 'low'.
evalMultiCategoryUse AI to categorize text into multiple categories from a comma-separated list. Returns array of category names. Optional quality parameter: 'high' (default), 'medium', 'low'.
evalStructuredUse AI to extract structured data from text using a Zod schema object. Returns { data } on success, or { error, data: null } on failure. IMPORTANT: Pass a Zod schema object created with z.object({...}), z.string(), z.number(), etc. JSON schema objects or plain object shapes will fail. When defining the schema, add helpful metadata via .description('...') on each field so the model knows the expected format or details (e.g. date formats, specific keywords). Optional quality parameter: 'high' (default), 'medium', 'low'.
generateGenerate free-form text using an LLM. Takes a user prompt and an optional system prompt to guide the response style/behavior. Optional quality parameter controls model intelligence: 'high' (default, most capable), 'medium' (balanced), 'low' (fastest). Returns an object { content: string, tokensUsed: number } with the generated text and provider usage telemetry.
querySearchSearch for information using LLM. Returns {content, results: [{url, title, snippet, source, publishedDate}]}
summarizeTextSummarize text content using LLM. Returns {summary, originalLength, summaryLength, tokensUsed}
synthesizeAnswerSynthesize research findings into a comprehensive answer using LLM. Returns {answer, question, sources, tokensUsed}
Write & mutating
0How install works
- 01
Choose a plan
One Workspace package. Your hosted workspace is prepared for you.
- 02
Install
Add LLM from the plugin directory — one click, no config files.
- 03
Connect & enable
Authorize LLM, then enable it for the agents that need it.
Put LLM in the hands of a persistent agent team.
Start a Workspace trial and install LLM in minutes.
Get started