Harnesses

Claude Code

Anthropic's CLI. Headless -p mode with a real advisor, JSONL transcripts per project, JSON settings at four scopes, MCP in ~/.claude.json.

getHarness: getHarness("claude")
binary: claude
invoke modes: 4 of 6 · official
model listing: no
mcp config: json · standard
agents file: ~/.claude/CLAUDE.md

Claude Code is the harness most of the others borrow from: Cursor, Copilot, OpenCode and Mastra Code read .claude/skills/, Grok loads CLAUDE.md. Its own layout is settings in JSON at user, project, local and managed scopes, transcripts under ~/.claude/projects/ in a directory named after the dash-encoded working directory, and MCP servers in ~/.claude.json rather than in the settings file, which surprises everyone once.

--tools "" removes every tool from the model, so the advisor mode is real. There is no native read-only sandbox, so readOnly: true is rejected rather than approximated with a permission prompt you'd never see in headless mode.

Capabilities and invocation

mcp: yesvision: yesaudio: novideo: notools: yesstreaming: yes

Binary claude, evidence official. Headless print mode; add --output-format json for structured output. Model selection appends --model {model}.

advisor

{}

-p {prompt} --tools

advisor · json

{ structured: true }

-p --output-format json {prompt} --tools

read-only

{ readOnly: true }

rejected

read-only · json

{ readOnly: true, structured: true }

rejected

agent

{ tools: true }

-p {prompt}

agent · json

{ tools: true, structured: true }

-p --output-format json {prompt}

Paths

Templates as the registry stores them. ~, ${HOME} and %VAR% expand in resolve(); entries tagged with a platform are dropped on the others. Open in the explorer to see them expanded for a home directory of your choice.

config

~/.claude/settings.jsonuser · official · all
~/.claude.jsonuser · official · allGlobal MCP server configurations.
.claude/settings.jsonproject · official · all
.claude/settings.local.jsonproject · official · allLocal overrides, typically gitignored.
.claude/.mcp.jsonproject · official · allProject-scoped MCP server configurations.
/etc/claude-code/managed-settings.jsonsystem · official · linuxEnterprise managed settings.
/Library/Application Support/ClaudeCode/managed-settings.jsonsystem · official · darwinEnterprise managed settings.

sessions

~/.claude/projects/<dash-encoded-cwd>/*.jsonldata · official · allPer-project session transcripts in JSONL format.
~/.claude/history.jsonldata · official · allGlobal prompt history across all projects.

instructions

CLAUDE.mdproject · official · all
.claude/CLAUDE.mdproject · official · allAlternative project-level location.
~/.claude/CLAUDE.mduser · official · allGlobal user-level instructions.

skills

.claude/skills/project · official · all
~/.claude/skills/user · official · allGlobal user-level skills.

commands

.claude/commands/project · official · allLegacy slash commands, merged into skills.
~/.claude/commands/user · official · allLegacy global slash commands.

hooks

.claude/hooks/project · official · all
~/.claude/hooks/user · official · allGlobal user-level hooks.

MCP servers

~/.claude.jsonuser · json · standard · mcpServers

.mcp.jsonproject · json · standard · mcpServers

Detection

envVars

CLAUDE_CODECLAUDECODECLAUDE_CODE_ENTRYPOINTCLAUDE_CONFIG_DIR

projectMarkers

.claude/settings.json.claude/settings.local.json.claude/skillsCLAUDE.md

Persistence

JSON - Settings and MCP config files. (official)
JSONL - Session transcripts and global history. (official)

syncAgentsFiles links ~/.claude/CLAUDE.md to the master bundle.

Edit this pageorReport an issue

Overview

Twelve harnesses, one page each, with every path, template and marker straight from the registry.

Codex CLI

OpenAI's CLI. codex exec with a native read-only sandbox, TOML config with comments preserved on edit, sessions as JSONL plus a SQLite state database.