Orchestrate teams of AI agents.Any model. One pipeline. At cost.
LoopLlama is the backend for agentic workflows. Describe the project, we assemble the agent crew, run the loop across whichever models fit each step — Claude, GPT, Gemini, Grok, DeepSeek, Mistral or your own — and return results. Model usage is billed at the provider's list price, with no markup.
Usage-based platform pricing · Models passed through at cost · See model rates
Everything you need to run agent teams in production.
LoopLlama is opinionated about the hard parts so your team can focus on the work.
Composable agent crews
Define each agent's role, prompt, model, tools, and token budget. We handle sequencing, persistence, and metering.
Stateful, durable workflows
Runs pause for human input or approval and resume exactly where they left off. They survive deploys and restarts — interrupted work is recovered automatically — and can run on a cron schedule.
Built-in connectors
13 out of the box — Salesforce, HubSpot, Gmail, Outlook, Slack, Twilio, Plaid, DocuSign, Sheets, Calendar, MySQL, Postgres, HTTP — with per-agent capability toggles and approval gates on writes.
Built-in observability
Per-step traces, token accounting, and replay. Debug a multi-agent run like a stack trace.
Any model, at cost
Route each step to Claude, GPT, Gemini, Grok, DeepSeek, Mistral or your own endpoint. Model usage is billed at the provider's list price — 0% markup.
Train your own models
Fine-tune GPT-4.1 or Mistral on uploaded examples — or on the steps your workflows have already produced — and route any agent to the result. Training and inference at the provider's list price.
Live web research, cited
Drop Perplexity-style search into any pipeline: quick answers, Pro Search, Deep Research reports and fact-checks on Sonar models, every claim with a source.
AI workflow builder
Describe the job in plain English. The builder inspects your connected Salesforce schema and drafts the crew — roles, prompts, tools and approval gates — ready to run or edit.
Templates & Agent Hub
Start from 34 ready-made lending workflows or 30 reusable agent roles — triage specialists, CRM updaters, renewal outreach — and plug them into any workflow.
Secure by default
Encrypted connector credentials, hashed API keys, role-based access, and an append-only audit log.
From brief to delivered artifact in one API.
No agent framework to learn. No prompt graphs to maintain. Just describe the work.
Describe the goal
Send a project brief — natural language, structured fields, or both. We parse it into a typed workflow spec.
We assemble the crew
Start from a template, pick roles from the Agent Hub, or let the AI builder draft the crew from your own CRM schema. Override any of it — models, tools, approval gates.
Run, monitor, deliver
Poll the run, or get a signed webhook when it completes, fails, or pauses for a human. Inspect every step's trace.
One call. A team of agents.
The LoopLlama API exposes workflows as a single endpoint. Trigger a run, then poll it or let a webhook tell you when it's done.
- Official TypeScript SDK, or plain REST from any language
- Signed webhooks for completion, failure, and human-input events
- Human-in-the-loop approvals over the same API
- Per-step traces with token counts on every run
import { LoopLlama } from "@loopllama/sdk";
const ll = new LoopLlama({ apiKey: process.env.LOOPLLAMA_API_KEY });
const workflow = await ll.workflows.create({
name: "Q3 brief",
crew: [
{ role: "planner", systemPrompt: "Outline the brief." },
{ role: "writer", systemPrompt: "Write it. Be concrete." },
{ role: "reviewer", systemPrompt: "Flag gaps and fix them." },
],
});
const run = await ll.runs.create(workflow.id, {
input: "Draft the Q3 product brief from these notes: ...",
});
const result = await ll.runs.poll(run.id);
console.log(result.output); // → markdown briefBuilt for the work your team is already doing.
The same API, pointed at six industries' back offices. See all solutions →
Lending operations
Triage broker submissions, chase stips, underwrite bank cash flow through Plaid, and draft approvals — with a human on every send.
Insurance agencies
Triage quote requests, chase applications and loss runs, work the 90-day renewal list, and issue certificates from the AMS.
Property management
Answer inquiries with real showing slots, chase applications, send leases, triage work orders, and follow up on late rent.
Recruiting & staffing
Screen applicants against the job order, schedule interviews, chase credentials, and send offer letters through DocuSign.
Accounting firms
Chase client documents, review a month of Plaid transactions down to the exceptions, and run the month-end checklist.
Legal intake
Qualify inquiries, run conflict checks against the matter database, send engagement letters, and draft client status updates.
Questions about running agent teams.
- What is LoopLlama?
- LoopLlama is an orchestration API for teams of AI agents. You define a workflow with a crew of agents, then trigger runs over a simple REST API — LoopLlama handles planning, sequencing, model calls, persistence, and usage metering.
- How does LoopLlama orchestrate multiple AI agents?
- Each workflow defines a crew of agents with distinct roles. On a run, LoopLlama executes the crew step by step — each agent sees the original input plus the prior agents' output — and records every step with its tokens and timing for replay and debugging.
- Which AI models does LoopLlama support?
- Any of them, in one pipeline. Each step of a workflow can run on a different model from a different provider — Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok, DeepSeek, Mistral, Perplexity Sonar for web-grounded cited answers, or any OpenAI-compatible endpoint you host — without rewriting the workflow. Tool calling, approvals and checkpoints work the same on every provider.
- What do AI models cost on LoopLlama?
- Exactly what the provider charges. Model usage is metered per token at each provider's published list price and passed through to you at cost — no markup, no bundled 'credits'. Your plan covers the platform; the model line on your invoice is the provider's price.
- Can I train or fine-tune my own model on LoopLlama?
- Yes. Launch a supervised fine-tuning job on GPT-4.1, GPT-4.1 mini, Mistral Small or Codestral from the dashboard or the API, using an uploaded JSONL file or the completed steps a workflow's agent has already produced. LoopLlama tracks the job and, when it finishes, the fine-tuned model is routable on any agent in your account — billed at the provider's training and inference list prices with no markup.
- Can agents use my own tools and data?
- Yes, through 13 built-in connectors: Salesforce, HubSpot, Gmail, Outlook, Slack, Twilio SMS, Plaid, DocuSign, Google Sheets, Google Calendar, MySQL, PostgreSQL, and any authenticated HTTP API. You enable specific capabilities per agent (read mail, query CRM, send a text), and can require human approval before any send or write. Connector credentials are encrypted at rest.
- Do I have to design the agent crew myself?
- No. Install one of 34 ready-made workflow templates, compose a crew from 30 reusable Agent Hub roles, or describe the job in plain English and the AI workflow builder drafts the crew for you — grounded in your connected Salesforce schema. Everything it produces is editable.
- What happens if a run is interrupted?
- Runs are durable. Every instance runs a worker that claims runs with a lease; if a deploy, crash, or restart interrupts one, another worker picks it up and resumes from the last completed step. A step that could have sent an email or written to a system is never replayed blindly — the run stops with a clear message instead.
- How does LoopLlama pricing work?
- Two transparent lines. The platform is usage-based: you pay only for the workflow steps your agents run. The free Developer tier includes 25 steps per month as a hard cap; paid plans include a larger monthly allotment and bill additional steps at a flat per-step overage rate. Model usage is passed through at the providers' list prices with zero markup.
- Can a workflow pause for human approval?
- Yes. Human-in-the-loop gates let an agent pause for input or approval on sensitive actions and resume exactly where it left off, so a person stays in control without stalling the run.
Stop wiring agents. Start shipping.
Get an API key, run your first workflow in under five minutes, and pay only for the steps your agents run — with every model billed at cost.