Skip to main content
Karta runs harness-native agent projects in production. Bring an agent defined to run in a supported harness: instructions, skills, tools, sub-agents, hooks, and config. Karta packages that project as a release, serves it through stable APIs and widgets, and creates durable kartas for the users, virtual employees, backend jobs, or fleet members that should each have their own workspace and memory. Use Karta when one agent project needs to create many agent instances: one per user for personal AI assistant use cases, one per team virtual employee, one singleton for a backend task, or one per fleet member. Karta adds release control, metering, budgets, webhooks, audit, and rollback around those instances.

Choose your path

Pick the right route for first deploys, widgets, API integrations, operations, and enterprise review.

Quickstart

Deploy a starter or existing agent project in under 10 minutes.

Core concepts

Agents, kartas, sessions, sub-agents, streaming, and releases.

Deploy

The ship, serve, consume loop, releases, and rollback.

API reference

/v1 sessions, streaming, and OpenAI- and Anthropic-shaped adapters.

​
Start by job

I want my first deploy

Install the CLI, scaffold an agent, deploy it, and open the hosted chat page.

I already have an agent project

Keep your harness-native files and add the Karta deploy manifest.

I need a chat widget

Add the hosted widget, then layer on identity, commands, and theming.

I need an API integration

Use sessions and messages directly, or choose an OpenAI or Anthropic adapter.

I need production review

Review architecture, security, spend controls, audit, and operations.

I need to operate agents

Manage releases, sessions, logs, budgets, webhooks, billing, and audit.

​
What Karta standardizes

One deploy creates a versioned release and serves it through stable consumer surfaces:
Each session resolves to the durable karta you choose: one per user, one per virtual employee that works with a team, one per backend job, or one per fleet member. That id decides which agent instance, workspace, and memory the session uses. See Kartas & memory.

​
Harness-defined agents

Teams usually choose one of two agent shapes. Framework-based agents use application code to control the workflow. LangGraph, CrewAI, AutoGen, and the OpenAI Agents SDK help programmers define state, routing, tool calls, and control flow. Harness-defined agents use natural-language instructions, skills, sub-agents, tools, and MCP configuration to give the model room to plan each turn dynamically. The harness runs that loop. Karta is built for this second shape: bring an agent defined for a supported harness, and Karta adds releases, durable kartas, sessions, identity, budgets, webhooks, and audit around it.

Framework-based agent

Programmer-controlled workflow: code owns state, routing, and the order of operations.

Harness-defined agent

Model-planned workflow: instructions and harness context guide the agent as it adapts to each task.

​
What you get

Multi-user sessions

Route conversations by metadata; multiple humans and AIs in one session, each message attributed to its sender.

Streaming-first

HTTP, CLI, SDK, and widget calls share typed events: text, tool use, reasoning, approval prompts, errors, and done.

Durable kartas

Persistent workspace and memory for each user-facing agent, virtual employee, backend job, or fleet member. Each karta is its own agent instance.

Metering & budgets

Per-org token and cost caps with precise usage tracking. Hit a cap and the API returns 402 before running anything.

BYOK

Bring your own provider key (Anthropic, OpenAI, Bedrock, Vertex, OpenRouter), encrypted at rest.

Deploy & rollback

Immutable versioned releases. Activation and rollback move the active release pointer.

​
Two planes

Karta separates running your agents from managing your account - the boundary security and platform teams should verify before launch:

Data plane - runs your agents

The request path: sessions, harness execution in isolated per-session sandboxes, release serving, streaming, and request-time budget enforcement.

Control plane - manages your account

The system of record: identity, team membership and roles, API keys, usage metering and budgets, billing, BYOK keys, outbound webhooks, and the audit log.
The split is the central design decision: the plane that runs agent code holds none of your money-and-identity state. A compromise of a running agent cannot by itself reach your billing, your team’s credentials, or another tenant’s data. See How Karta works.

​
Who Karta is for

Agent developers

Author an agent project in your own repo and harness, then deploy it on Karta using the dashboard, the karta CLI, the SDKs, and your own backend.

Product engineers

Embed agents in your app with the hosted widget, a custom UI, or an API integration that keeps browser credentials short-lived and scoped.

Platform architects

Review the trust boundary, isolation model, identity, spend gates, releases, webhooks, audit, and operational controls before launch.

DevOps

Deploy, configure, and operate agents on Karta - releases, rollbacks, keys, webhooks, and live-session debugging.

Admins

Track usage and cost, set budgets and thresholds, and manage billing for the agents you run.
Your end users can reach the agent through your hosted widget, your custom UI, or your backend. Browsers receive publishable widget keys or short-lived session tokens, never long-lived API keys.

​
Where to go next

Choose your path

Route yourself by job: first agent, existing project, widget, API, ops, or architecture review.

Build your first response

Install, scaffold or select a folder, deploy, and open the hosted page.

Understand the model

The vocabulary the rest of the docs assume.

Author an agent

Agent layout, sub-agents, skills, and karta.toml.

Run a support bot locally

A complete end-to-end walkthrough.

Production readiness

Launch checklist for access, spend controls, releases, webhooks, audit, and support.

Production architecture review

The enterprise path through security, tenancy, credentials, spend controls, webhooks, and audit.