OPEN SOURCE · LAUNCHING JULY 2026

Give your app an agent that does the work.

Sogentic AI is an embeddable runtime for building agents that take real actions inside your product. Imagine Claude Code or Codex for your app, and friendly enough for your users. You describe what the agent can do in plain declarations. The runtime handles the hard part.

1 primitive: the module
5 plain-English declarations
0 rewrites of your business logic
What it is

A runtime you embed, not a platform you migrate to.

Your application supplies four things in plain English: the tools an agent can use, the context it knows, the prompts that shape how it behaves, and the relationships between agents. The runtime does everything else: running the loop, picking tools, assembling context, spawning sub-agents, and surfacing every action so users can see, edit, and undo it.

Your application: you own this
Host App
UI · Database · Auth · Tool implementations · State
supplies plain declarations
Sogentic runtime: we handle this
Agent Runtime
Orchestration · Context assembly · Permissions · Sub-agent loops · Visibility & undo
The shift

If you can build software, you can now build agentic systems.

There's a gap in the industry right now. Plenty of engineers can build great software, but building a reliable agentic system has meant wrestling with execution loops, context windows, tool routing, retries, and orchestration plumbing. That's a different skill set, and it's kept most teams out. Sogentic closes that gap. The architecture is still yours to design; that's a real skill. But the technical machinery underneath is no longer your problem.

Building an agent today

You become an infrastructure engineer first.

  • Hand-write the execution loop and manage every state transition
  • Babysit the context window as it bloats and drifts
  • Wire up tool routing, retries, and error recovery yourself
  • Reinvent permissions, confirmation, and undo from scratch
  • Glue multi-agent handoffs together by hand
Building an agent with Sogentic

You stay an architect. You declare; we run.

  • Describe each tool and what it does in plain English
  • Context stays sharp automatically, rebuilt fresh each turn
  • Tool routing, retries, and self-correction are built in
  • Permissions, confirmation, and undo come for free
  • Multi-agent systems are just modules referencing modules

The hard part was never the plumbing. It was knowing what to build. Now that's the only part left.

Two ways to use it

Agents that talk. Agents that run on their own.

The same runtime powers both a conversational copilot and a headless background worker. One primitive, two very different products.

USE CASE 01
Conversational · user-facing

An AI your users talk to.

Drop a copilot into your product that takes action on the user's behalf. Simple requests run in one step. Complex ones get a visible plan the agent works through. It can ask clarifying questions, chain multiple actions, and explain itself as it goes.

Every action is visible before it runs. Users can edit a step inline, or undo one with a click. Think of the power of Claude Code or Codex, delivered to your users in a way that feels safe and friendly.

"Find every overdue invoice from last quarter, group them by client, and draft a reminder for each." And it just does it, showing every step.
USE CASE 02
Headless · autonomous

Agentic workflows, no chat required.

Not every agent needs a person typing to it. Trigger a module from an event in your app or on a timer, and let it carry out a multi-step job autonomously, with the same tools, context, and safety guarantees.

A new record appears, a threshold is crossed, a schedule fires; the agent wakes up, reasons about what's needed, takes the right actions, and logs everything for review. Agentic automation that runs while your users sleep.

Every morning at 7am: pull yesterday's signups, score each lead, enrich the promising ones, and queue a personalized follow-up, untouched by human hands.
Why it holds up in production

Built for real apps, not demos.

Safety by structure

An agent can't misuse a tool it was never given. Permissions strip tools out before the model can even see them; restraint is built into the shape, not bolted on with prompts.

See, edit, undo

Every action streams to your UI before it commits. Users can correct a payload inline or revert a completed action with one click. Undo logic lives right on the tool.

Sharp context, always

Instead of dragging an ever-growing history, the runtime rebuilds context fresh each turn, pulling only what's relevant. Lower cost, tighter focus, no drift.

One primitive, any scale

A tiny text-formatter and a sprawling multi-agent coordinator share the exact same shape. Compose agents by having modules reference modules. Nothing new to learn.

Fails gracefully

Tools fail with messages the agent can act on, so it self-corrects instead of stalling. When it truly can't proceed, it asks; it never guesses silently.

Embeddable anywhere

An unopinionated package: no database to adopt, no platform to deploy to. Run it in a web server, a CLI, or a background queue. Your stack stays your stack.

How it compares

A different shape than the frameworks you know.

LangGraph, the Vercel AI SDK, and LlamaIndex are excellent at what they're built for. Sogentic is built for something they're not: cleanly embedding an action-taking agent into an existing product, with enterprise-grade visibility and control.

Capability Sogentic AIembeddable runtime LangGraphstate-graph engine Vercel AI SDKfrontend-first LlamaIndexRAG-first
Core primitive One module: same shape for a single agent or a whole coordinator State graph: nodes and routing edges you draw in code A tool-loop agent wired into web primitives Indices and query engines, built around retrieval
How you configure it Plain declarations: describe tools, context, prompts, relations Code-first: explicitly build the graph by hand Functional hooks and callbacks in code Class instantiation with heavy data-loader abstraction
Context handling Rebuilt fresh each turn: only what's relevant, no drift Accumulates everything in a growing state array A standard, ever-growing message history array Vector search injects chunks before each call
Permissions Structural: unusable tools are invisible to the model Prompt or custom-node guardrails Intercepted at runtime in your app code Not a primary concern
Before an action runs Live edit: users can change the action inline, then continue Execution pauses; you update state via API A yes/no approval hook Not a primary concern
After an action runs One-click undo: reversal logic defined on the tool "Time travel": rewind graph state to a checkpoint You catch and reverse it manually Not a primary concern
Runs headless Yes: chat, CLI, worker, or timer, same runtime Yes, but inside its own state server Hard: built around the front-end stream Yes, for retrieval tasks

Reach for LangGraph

when…

You're running a long, crash-recoverable background pipeline with thousands of non-linear transitions, closer to a workflow engine than a copilot.

Reach for Vercel AI SDK

when…

You're a frontend developer who wants a standard B2C chatbot streaming straight into a React view, with near-zero setup.

Reach for Sogentic

when…

You have a real product and want to add an agent that performs multi-step actions for users with strict permissions, full visibility, and undo, without rewriting your business logic.

Open source

Free, open, and yours to build on.

Sogentic AI is open source. Read the code, fork it, extend it, ship it in production, no strings. We believe the agent runtime that powers your product should be something you can fully see and own.

Public release · July 2026
Don't want to build it yourself?

Or let us do it for you.

The runtime is free and open. But if you'd rather not wire it up yourself, our managed service builds and hosts your agents for you: designed, integrated, and run end-to-end.

  • We design and build your agents against your product and goals.
  • We host and operate the runtime so you don't touch infrastructure.
  • We optimize it to make your agents faster and cheaper to run.
  • Simple, transparent pricing scored in points: we rate the difficulty of what you need, and the price follows.
How pricing works
1 pt = $100

We score the difficulty of a build in points. Points map directly to a one-time setup fee, plus a monthly service fee and API costs.

A 10-point build →
$1,000 setup fee + monthly service fee + API costs.