Open Interpreter reads its configuration from ~/.openinterpreter/config.toml. You can also drop a .openinterpreter/config.toml into a project to layer project-specific settings on top.

Where settings come from

Settings layer in this order, with later layers winning:

To see exactly where each value came from:

/debug-config

The settings you reach for most

# Default model and provider
model = "gpt-5-codex"

# When to ask before running things: "untrusted" | "on-request" | "never"
approval_policy = "on-request"

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

# Reasoning effort: "low" | "medium" | "high" | "none"
model_reasoning_effort = "medium"

# Communication style
personality = "concise"

# Where the agent writes logs
log_dir = "~/.openinterpreter/log"

Profiles

Profiles are named groups of overrides for different contexts.

[profiles.fast]
model = "gpt-5-codex-mini"
model_reasoning_effort = "low"

[profiles.review]
model = "gpt-5-codex"
model_reasoning_effort = "high"
sandbox_mode = "read-only"

Use one with:

interpreter --profile review

Override values from the CLI

-c takes a TOML expression and applies it to the active config:

interpreter -c model='"gpt-5-codex-mini"' -c approval_policy='"never"'

Useful for one-off runs and scripts.

MCP servers

Open Interpreter can connect to Model Context Protocol servers. Define them under [mcp_servers]:

[mcp_servers.docs]
command = "docs-server"

By default, MCP tools run one at a time. To allow parallel calls for a server whose tools are safe to run together:

[mcp_servers.docs]
command = "docs-server"
supports_parallel_tool_calls = true

MCP tool approvals

Set a default approval mode for every tool exposed by a server, and override individual tools:

[mcp_servers.docs]
command = "docs-server"
default_tools_approval_mode = "approve"

[mcp_servers.docs.tools.search]
approval_mode = "prompt"

See the MCP servers guideMCP servers guidePreview this page. for the full picture.

Feature flags

Optional behavior lives behind [features].

[features]
apps = true            # ChatGPT connector surface ($ mentions, /apps)
plugins = true         # Plugin bundles for skills, MCP, connectors
child_agents_md = true # Hierarchical AGENTS.md guidance

Apps and plugins stay off by default so the base config stays provider-neutral.

Notify hook

Open Interpreter can run a shell hook each time a turn finishes. Wire it up under [notify]. The notification payload includes a client field identifying the surface that started the turn (the TUI reports interpreter-tui).

Custom CA certificates

Behind a corporate proxy that intercepts TLS? Point Open Interpreter at your bundle:

export CODEX_CA_CERTIFICATE=/etc/ssl/corp-bundle.pem

If CODEX_CA_CERTIFICATE is not set, Open Interpreter falls back to SSL_CERT_FILE. If neither is set, it uses your system root certificates.

The PEM file may contain multiple certificates. OpenSSL TRUSTED CERTIFICATE labels are tolerated, and well-formed X509 CRL sections are ignored.

SQLite state

Sessions and memory are stored in SQLite under sqlite_home. Override it with the config key or CODEX_SQLITE_HOME. By default, workspace-write sessions store under a temp dir, while other sandboxes use ~/.openinterpreter.

Plan mode defaults

plan_mode_reasoning_effort lets you override reasoning effort while in Plan mode. Set it to "none" to explicitly disable reasoning when planning. Leave it unset to inherit the Plan preset default (medium).

Notices

The [notice] table stores "do not show again" flags for some UI prompts. Delete entries here if you want a notice to appear again.

JSON Schema

A generated JSON Schema for config.toml lives at codex-rs/core/config.schema.json in the source tree. Useful for editor autocomplete or validation in CI.

Quitting hint

Pressing Ctrl+C once shows a one-second hint (ctrl + c again to quit). The second press exits. This catches accidental quits without making intentional exits slow.