Skip to main content
Karta has a small, consistent vocabulary. Learn these eight terms and every other page becomes easier to read. If you are here to solve a specific job, use Choose your path first; come back here when a page uses a term you want to pin down.

Agent harness

The runtime that executes the agent loop. See Supported harnesses for the harness formats Karta detects and runs. The harness owns the model loop, tools, MCP connections, memory, context, and turns.

Agent project

The deployable unit: a folder that configures a harness (CLAUDE.md or AGENTS.md, a native harness directory, sub-agents, skills, and karta.toml). The CLI and API use this term for the same object.

Karta

A durable agent instance created from one agent project: one per user for a personal AI assistant, one per team virtual employee, one singleton for a backend task, or one per fleet member. Each karta has its own workspace and sessions.

Session

A lightweight orchestration handle - metadata, participants, current sub-agent, pending approvals. The session record is not the message store; conversation history lives with the harness.

Participant

A human or AI actor in a session. Many can share one session; each message is attributed to its sender.

Sub-agent

An optional harness-native role inside one agent project, such as a .claude/agents/*.md file. It is packaged with the parent agent release, under the same agent endpoint.

Release

An immutable, versioned snapshot of an agent, materialized from a published artifact. Activation is a pointer flip.

Runtime and management boundary

Karta separates agent execution from account management. Running agent code does not become billing, key-management, or audit authority.

​
How they fit together

A request carries a credential, Karta validates it and checks the budget, resolves (or creates) the session and karta, hands the turn to the harness, and relays the harness’s typed event stream back to the caller. Usage is metered back to the account afterward. Walk through it in How Karta works.

​
Who owns what

The harness owns the model loop, tools, MCP, memory, context, and conversation history inside a running karta. Karta owns the platform around it: durable karta identity, sessions, participants, multi-tenancy, releases, metering, and the uniform API. Because your agent stays in the harness’s native format, Karta can read the project layout directly.

Choose your path

Route yourself by job before going deep on the concept model.

How Karta works

Per-session isolation, the request path, and the runtime/management boundary.

Agents

What you actually ship, and how a harness is detected.

Sessions & participants

Multi-turn state, multi-party conversations, and attribution.

Streaming events

The typed event model every surface is built on.

Kartas & memory

Durable agent instances, memory, privacy, and how to choose the id that selects them.