Supported harnesses

This release supports five harnesses. Each one is driven through its own programmatic interface, never by scraping a terminal

from enum import Enum

class Harness(Enum):
    CLAUDE_CODE = "claude_code"
    CODEX = "codex"
    OPENCODE = "opencode"
    DEEPAGENTS = "deepagents"
    TOOL_LOOP = "tool_loop"

Harness is a plain Enum. A StrEnum member would compare equal to its string and let "codex" through, so litellm.agent() and the other entry points check isinstance(harness, Harness) and raise TypeError otherwise.

Capabilities

Capability	Claude Code	Codex	OpenCode	Deep Agents	Tool Loop
Any gateway model group	yes	yes	yes	yes	yes
Cost tracking	yes	yes	yes	yes	yes
Skills	yes	yes	yes	yes	no
Structured output	yes	yes	yes	yes	yes
Detach and resume	yes	yes	yes	yes	no
Built-in tool filtering (`disable_tools=`)	yes	no	yes	yes	no
Custom Python tools (`tools=`)	no	no	no	yes	yes
History (`s.history()`)	no	no	no	yes	yes
Permission modes	`read-only`, `edit`, `full`	`read-only`, `full`	`read-only`, `edit`, `full`	`read-only`, `edit`, `full`	`ask`, `full`

permissions="ask" with interactive approval isn't available on the CLI harnesses in this release. Asking a harness for something it doesn't support raises CapabilityUnsupported before the runtime starts, and it never falls back to a looser mode. You can read the same table in code.

caps = litellm.agent_capabilities(Harness.CODEX)
caps.tool_filtering     # False
caps.permission_modes   # frozenset({'read-only', 'full'})

How each one is driven

Harness	Driven through	Speaks to its model	Runs in	Needs in sandbox
`CLAUDE_CODE`	`claude -p --output-format stream-json`	Anthropic Messages	sandbox	`claude`
`CODEX`	`codex exec --json`	OpenAI Responses	sandbox	`codex`
`OPENCODE`	`opencode run --format json`	OpenAI Chat Completions	sandbox	`opencode`
`DEEPAGENTS`	`deepagents` Python API	LangChain chat model	your process	nothing
`TOOL_LOOP`	LiteLLM tool loop	OpenAI Chat Completions	your process	nothing

The runtime binaries aren't installed for you in this release. Put them in your sandbox image, or on your PATH for sandbox.local. If the binary is missing, the call raises HarnessInstallFailed naming it.

Every harness takes a typed options class for settings that only make sense for that runtime. Passing another harness's options raises OptionsMismatch.

from litellm import Harness, CodexOptions

litellm.agent(
    Harness.CODEX, task, sandbox=box,
    options=CodexOptions(reasoning_effort="high"),
)

Capabilities​

How each one is driven​

Capabilities

How each one is driven