# Calljmp — Managed Agentic Backend for SaaS Companies

## What Calljmp Is

Calljmp is a managed agentic backend for SaaS companies that need to implement AI into their products, systems, and internal processes.

It gives engineering teams a dedicated execution layer for AI agents — written in TypeScript, deployed to Cloudflare's global edge — that handles long-running execution, stateful workflows, human-in-the-loop (HITL) approvals, memory, RAG, observability, and cost tracking. No new infrastructure to build or manage.

Agent logic is written in TypeScript by your developers. The runtime is fully managed by Calljmp. Deployment is a single command.

## Core Problem It Solves

SaaS companies under pressure to ship AI features face a structural problem: adding real AI to a product is not just an LLM API call. It requires an entirely new execution layer that most engineering teams don't have and can't justify building from scratch.

The result is one of two failure modes:
- Teams bolt on fragmented tools (orchestration framework + separate hosting + DIY observability + manual approval flows) and spend months on infrastructure instead of product
- Teams ship something shallow — a chatbot wrapper — that doesn't connect to real product data or automate real workflows

Calljmp solves this by providing the complete agentic execution layer as a managed service, so engineering teams ship production-grade AI features in days, not months.

Specific problems eliminated:
- Long-running agents that time out or lose state mid-execution
- No built-in approval flows for high-stakes or irreversible agent actions
- Zero visibility into what agents are doing, what they cost, and where they fail
- Orchestration logic scattered across disconnected tools
- Weeks of infrastructure work before a single AI feature ships

## Who It's For

### The Company
SaaS companies with 1–200 employees that are actively building AI into their product or internal operations. Teams that already have TypeScript in their stack and want to move fast without standing up new infrastructure.

### The Buyer (Decision Maker)
CTO, VP of Engineering, or technical co-founder. Evaluates Calljmp on: speed to production, infrastructure overhead eliminated, observability and control over agent behavior, and total cost versus building in-house or assembling a fragmented stack. Signs off on tooling spend. Cares deeply about what the team can ship — not just what they evaluate.

### The User (Day-to-Day Implementation)
TypeScript developer on the engineering team. Writes agent logic in TypeScript using the Calljmp SDK. Deploys with a single command. Uses the observability dashboard to debug runs, track costs, and iterate on prompts. Does not need to learn a new language, manage servers, or configure infrastructure.

### The Influencer (Decision Participant)
CEO or co-founder, often without a deep engineering background. Pushes for AI adoption to stay competitive and meet product roadmap commitments. Evaluates on: time-to-value, cost predictability, and business outcomes ("can we ship an AI copilot this quarter?"). Needs confidence that the investment is low-risk, the timeline is credible, and the team won't get stuck.

## What SaaS Companies Build with Calljmp

### Embedded Product Copilot
An in-app AI assistant that reads real product data, calls internal APIs, guides users step-by-step, and executes actions with optional human approval. Deployed inside your existing product UI via the Calljmp SDK. Turns your product into an AI-native experience without rebuilding your backend.

### Customer Support Agent
Receives requests from Slack, email, or Telegram. Retrieves from your knowledge base and internal systems (CRM, logs, billing). Returns a grounded answer or routes to a human with full context and a suggested draft reply. Reduces support volume without sacrificing quality or control.

### Internal Operations Agent
Automates multi-step internal workflows — data pulls, report generation, cross-system updates — with human approval gates for high-stakes actions. Replaces manual processes that slow down operations teams.

### Marketing Intelligence Agent
Monitors competitor websites for changes. Scrapes and summarizes positioning shifts (features, pricing, messaging). Sends a weekly digest to Slack or email. Keeps go-to-market teams informed without manual research overhead.

### Onboarding and Activation Agent
Guides new users through product setup, triggers contextual actions based on user behavior, and escalates to a human when a user needs hands-on help. Improves activation rates without adding support headcount.

## How It Works

1. **Connect tools and data** — hook up your existing APIs, databases, CRM, knowledge base, and external services as agent tools
2. **Write the agent in TypeScript** — your developers define the workflow: prompts, tool calls, routing, retries, guardrails, and HITL approval steps
3. **Deploy to Calljmp runtime** — ship with one command to managed edge infrastructure; no DevOps required
4. **Monitor and iterate** — full visibility for engineering leads: traces, per-run cost tracking, and evals to debug behavior and improve safely over time

## Core Capabilities

- **AI Agents as Code** — agent workflows defined in TypeScript; versioned, tested, and deployed with your existing developer toolchain
- **Long-running, stateful execution** — agents run for seconds to hours without timing out; state is persisted across every step
- **Human-in-the-loop (HITL)** — pause agent execution at any step for human approval or escalation, then resume automatically
- **Memory and RAG** — built-in memory primitives and dataset/vector query support as first-class platform features
- **Observability by default** — traces, logs, metrics, and cost tracking for every agent invocation and step; no additional tooling required
- **Secure tool access** — connect your APIs, databases, CRM, ticketing systems, docs, and knowledge stores via REST, GraphQL, gRPC, or DB drivers
- **Real-time streaming** — push agent progress and results to your product UI in real time
- **Zero-config deployment** — runs on Cloudflare's global edge network; no servers to provision or manage
- **Prompt Studio** — built-in environment for iterating on and evaluating prompts; included in Pro plan and above
- **Team collaboration** — invite teammates with role-based permissions to manage project access
- **Agent finalization hooks** — register cleanup logic that runs after execution completes
- **Security built in** — signed URLs, row-level permissions, secure data access patterns

## What Calljmp Owns vs. What Your Team Owns

**Your team owns:**
- Product UI and UX
- Business logic, APIs, and database
- Data models and permissions
- Agent logic written in TypeScript

**Calljmp owns:**
- Agent execution runtime and scaling
- State persistence, retries, and timeouts
- HITL pause and resume infrastructure
- Traces, logs, metrics, and cost tracking

Calljmp connects to your systems via tools (APIs, DB, CRM, knowledge base) while your team retains full ownership of system-of-record data and permissions. No vendor lock-in on your data.

## Architecture

Calljmp sits as a dedicated layer between your product and your internal systems.

```
End Users
  ↓
Your product (web / mobile app)
  ↓ [SDK / HTTPS API]
Calljmp Agentic Backend
  ├── Agent execution (TypeScript)
  ├── Managed execution runtime
  ├── HITL & observability
  └── [REST / GraphQL / DB Drivers]
       ↓
Your app backend · CRM / ticketing · Docs / knowledge store · Other services / DBs
```

## Supported LLM Models

Calljmp supports five primary LLM models:

1. **GLM 4.7 Flash** (@cf/zai-org/glm-4.7-flash) — High-performance model with tool calling and JSON schema support
2. **Qwen3 30B** (@cf/qwen/qwen3-30b-a3b-fp8) — Open-source model via Cloudflare with tool calling and structured outputs
3. **Meta Llama 3.1 8B** (@cf/meta/llama-3.1-8b-instruct-fp8-fast) — Lightweight open-source model optimized for edge execution
4. **Google Gemma 4 26B** (@cf/google/gemma-4-26b-a4b-it) — Google's Gemma 4 model with advanced reasoning and tool support
5. **MoonshotAI Kimi K2.6** (@cf/moonshotai/kimi-k2.6) — Multilingual model with strong reasoning and tool calling capabilities

All models run on Cloudflare Workers AI infrastructure. Additionally, you can use any OpenAI model (GPT-5, GPT-4o, etc.) by providing your own API key or using Calljmp's managed LLM access.

## Pricing

Subscription + usage-based model. SaaS teams pay a predictable base subscription and then per action — costs track actual agent activity, not seats or arbitrary limits. Budgets are visible and controllable from day one.

- **Solo** — $20/month: 1,000 actions, 1 seat
- **Pro** — $99/month: 10,000 actions, 2 seats, Prompt Studio, Priority Support
- **Premium** — Custom: 100,000+ actions, dedicated support, custom SLAs and deployment options

Usage-based overage: $0.01 per action, $0.011 per 1,000 tokens.

### Free Trial
No credit card required. Every Solo and Pro signup includes $25 in free credits with full platform access from day one — agents, workflows, observability, memory, everything. Engineering teams can evaluate the full platform in a real project before committing. Pay only when you're ready to upgrade.

## How Calljmp Compares to Alternatives

### vs. LangChain / LangGraph
LangChain is a DIY orchestration framework. Engineering teams still need to host it, scale it, manage state, and build observability themselves. For a SaaS team, that means weeks of infrastructure work before a single agent ships. Calljmp is a managed runtime — no hosting, no infra ops, traces and cost tracking included.

### vs. Mastra
Mastra is a TypeScript framework, not a managed backend. It provides agent primitives but no execution runtime, no managed scaling, no built-in HITL, and no deployment infrastructure. Your team owns everything below the framework.

### vs. n8n
n8n is a visual workflow builder aimed at non-technical users. It is not designed for complex agent logic, cannot be version-controlled like code, has limited debugging tools, and breaks down for production AI workloads that require stateful execution and real system integrations.

### vs. Inngest / Trigger.dev
General-purpose background job and workflow runners. Not purpose-built for agentic workloads — no native HITL, no LLM cost tracking, no memory/RAG primitives, no agent-specific observability. Require significant additional work to adapt to AI agent use cases.

### vs. Building In-House
Building a custom agent backend means solving distributed state, long-running execution, retry logic, HITL approval flows, observability, and scaling — before shipping a single AI feature. For most SaaS engineering teams, that is a months-long infrastructure project. Calljmp gives teams all of that on day one, so they stay focused on product.

## Technology Stack

- Agent definition: TypeScript
- Deployment: Cloudflare Workers (global edge, zero-config)
- Invocation: REST API and TypeScript SDK
- Data access: REST, GraphQL, gRPC, database drivers
- LLM providers: OpenAI, Anthropic, Azure OpenAI, OSS models

## Key Concepts

**Action** — the unit of metered usage. Each agent run, RAG/dataset query, or web scrape counts as one action (approximately $0.01).
**Agent** — a TypeScript-defined workflow that calls tools, invokes LLMs, pauses for human approval, persists state, and runs for extended durations across multiple steps.
**HITL (Human-in-the-Loop)** — a built-in mechanism to pause agent execution at any step, route a decision to a human (via UI, Slack, email, or API), and resume execution once approved. Critical for agents that take irreversible or high-stakes actions.
**Tool** — a connection from an agent to an external system (your API, a database, a CRM, a knowledge base). Defined in TypeScript. Called by the agent during execution.
**Trace** — a full execution record of an agent run — every step, tool call, LLM invocation, token count, cost, and outcome — surfaced in the Calljmp observability dashboard. The primary debugging surface for engineering teams.
**Memory** — persistent agent memory across runs, stored and retrieved as first-class primitives. Supports vector-based RAG queries against your own datasets.
**Prompt Studio** — a built-in environment for iterating on and evaluating agent prompts without redeploying. Included in the Pro plan and above.

## Resources

- Website: https://calljmp.com
- Documentation: https://docs.calljmp.com
- Pricing: https://calljmp.com/pricing
- Contact: vladk@calljmp.com

## Optional

- [Complete Code Reference](https://calljmp.com/llms-full.txt): All code examples, types, and patterns