Open Interpreter reads durable settings from TOML files. User-level config lives at:

~/.openinterpreter/config.toml

Project config can live in a trusted project under:

.openinterpreter/config.toml

Command-line overrides use -c key=value and apply only to that invocation.

Precedence

Higher-precedence values override lower-precedence values:

1. Built-in defaults
2. System or managed configuration
3. User config
4. Trusted project config
5. Selected profile
6. CLI overrides from -c, --enable, --disable, or dedicated flags

Use /debug-config in the TUI to inspect the effective values and where they came from.

Common Settings

model = "gpt-5.1-codex"
model_provider = "openai"

# "minimal" | "low" | "medium" | "high" | "xhigh"
model_reasoning_effort = "medium"

# "auto" | "concise" | "detailed" | "none"
model_reasoning_summary = "auto"

# "read-only" | "workspace-write" | "danger-full-access"
sandbox_mode = "workspace-write"

# "untrusted" | "on-request" | "never"
approval_policy = "on-request"

# "friendly" | "pragmatic" | "none"
personality = "pragmatic"

web_search = "cached"
log_dir = "~/.openinterpreter/log"

Profiles

Profiles are named groups of settings:

[profiles.fast]
model = "gpt-5.1-codex-mini"
model_reasoning_effort = "low"

[profiles.review]
model = "gpt-5.1-codex"
model_reasoning_effort = "high"
sandbox_mode = "read-only"

Use one with:

interpreter --profile review

CLI Overrides

-c accepts TOML-like values. Quote strings so your shell does not strip them:

interpreter -c model='"gpt-5.1-codex-mini"' -c approval_policy='"never"'

Feature flags also have short forms:

interpreter --enable hooks --disable memories

Feature Flags

Optional behavior lives under [features].

[features]
hooks = true
multi_agent = true
shell_tool = true
shell_snapshot = true
unified_exec = true
memories = false
apps = false
plugins = false
undo = false

Use /experimental in the TUI for interactive toggles when available.

Model Providers

Built-in providers are selected by model_provider. Custom OpenAI-compatible providers can be added under [model_providers.<id>]:

model_provider = "acme"
model = "acme-coder-large"

[model_providers.acme]
name = "Acme"
base_url = "https://api.acme.example/v1"
env_key = "ACME_API_KEY"
wire_api = "responses"

Provider credentials should usually come from environment variables or the credential store rather than inline tokens.

Harness

Open Interpreter adds a harness setting for compatibility modes that shape the agent surface like another coding harness while still running through the native Open Interpreter runtime.

harness = "kimi-cli"
harness_guidance = true

Supported values are implementation-dependent, but the current codebase includes native, Claude Code, DeepSeek TUI, Kimi CLI, Qwen Code, SWE-agent, and minimal harness modes. Use this when you intentionally need a harness-shaped prompt/tool surface. Leave it unset for normal Open Interpreter behavior.

harness_guidance lets Open Interpreter include a small reliability guidance block where that harness mode allows it. Set it to false if you need stricter harness behavior.

MCP Servers

MCP servers are configured under [mcp_servers]:

[mcp_servers.docs]
command = "npx"
args = ["-y", "@acme/docs-mcp"]
env = { ACME_TOKEN = "env:ACME_TOKEN" }
default_tools_approval_mode = "prompt"

Streamable HTTP servers use url:

[mcp_servers.search]
url = "https://mcp.example.com"
bearer_token_env_var = "MCP_TOKEN"

See MCP for transport, OAuth, and per-tool approval details.

Shell Environment

Use shell_environment_policy to control what environment variables are passed to spawned commands:

[shell_environment_policy]
inherit = "all"
ignore_default_excludes = false
exclude = ["AWS_SECRET_ACCESS_KEY", "DATABASE_URL"]
set = { CI = "1" }

History and Memory

Session history is stored locally. You can disable transcript persistence:

[history]
persistence = "none"

Memories are a separate experimental feature:

[features]
memories = true

[memories]
use_memories = true
generate_memories = true

See Memories.

Config Schema

The source tree includes a generated JSON Schema at:

codex-rs/core/config.schema.json

Use it for editor completion or CI validation when maintaining shared config.