Karta has a small, consistent vocabulary. Learn these eight terms and every other page reads easily.Documentation Index
Fetch the complete documentation index at: https://docs.karta.sh/llms.txt
Use this file to discover all available pages before exploring further.
Agent harness
The agent runtime — Claude Code, OpenCode, deepagents. It owns tools,
MCP, memory, context, and the agentic loop. Karta does not reinvent it.
Harness application
The deployable unit: a folder that configures a harness (
.claude/ +
CLAUDE.md, agents, skills). In code and the API it’s called a project.Session
A lightweight orchestration handle — metadata, participants, current agent,
pending approvals. It holds no messages; the harness does.
Participant
A human or AI actor in a session. Many can share one session; each message
is attributed to its sender.
Agent
A thin wrapper over an agent definition discovered from the harness
(
.claude/agents/*.md). Sessions route between agents and hand off.Gateway
Event ingestion and fan-out. Delivers messages to all participants —
in-process, or across Karta instances over HTTP.
Release
An immutable, versioned snapshot of a harness application, materialized
from a published artifact. Activation is a pointer flip.
The two planes
The data plane runs agent code; the control plane holds identity,
money, and audit. Separated by a trust boundary.
How they fit together
The mental model that matters most
Karta delegates; it does not duplicate. The harness is the single source of truth for conversation history, persistence, resumption, and tool/MCP integration. Karta keeps no second copy — which is why there are no sync races and why an example from the harness’s own docs runs unchanged inside Karta. Everything else — sessions, participants, the gateway, multi-tenancy, releases — is the platform Karta wraps around that delegated core.Read next
Architecture
The two planes, the trust boundary, and the request lifecycle.
Harness applications
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.