Skip to main content
Karta runs agents that are defined for a supported harness. The agent stays in that harness’s native project layout; Karta detects the layout, packages it as a release, and runs each turn through that harness. Use this page to choose the --harness value for karta create, or to check whether an existing agent project can be brought to Karta without changing its file format.

​
Current harnesses

For current version availability and exposed capabilities, use the public dashboard catalog at /harnesses.

​
Create a starter

karta create writes the native instruction file and marker directory for the selected harness. If you already have an agent project, keep its existing layout and run karta setup --enable from that folder.

​
Detection rules

Karta detects the harness from the folder layout. If a folder has markers for more than one harness, declare the harness explicitly in karta.toml:
Mixed markers are allowed when you declare the harness, but they are harder for teammates and CI to reason about. Prefer one harness layout per agent project.

​
What changes on Karta

Karta does not translate your agent into a Karta-specific agent format. The harness still owns the agent loop, tools, MCP connections, context, memory format, skills, sub-agents, and conversation history. Karta adds release packaging, durable kartas, sessions, identity, isolation, metering, budgets, webhooks, audit, and API surfaces around that harness-native project.

Agent structure

See the file layout for each supported harness.

karta.toml

Declare the harness and deploy settings when auto-detection is not enough.