Agent Architecture

Open Interpreter

Open Interpreter is an open-source terminal coding agent for local and open models. The current interpreter (also installed as i) is a Rust fork of OpenAI's Codex, so its native config is Codex-shaped: a TOML config.toml under an isolated home, $INTERPRETER_HOME (default ~/.openinterpreter), never $CODEX_HOME.

Source reviewed #40Checked 2026-09-07

Setup guide Source

Run type

CLI

Adapter

json-stdio

Source

OSS · openinterpreter/openinterpreter

Stars

★ 68k

Archive Context

This page is one host entry inside the full architecture archive and keeps multi-surface products visible instead of collapsing them into a single run type.

42 host archive

Host page

open-interpreter

Run cohort

CLI: 30

Archive route

/agents

Host Surface Perspectives

The host is the primary architecture boundary. Each runnable surface below separates what the host owns from what agent-connector can safely install for Open Interpreter.

1 surface entry

CLI

cli

Host entry: A terminal command owns the process, current working directory, stdin/stdout, shell environment, and project file discovery.
Host owns: The CLI host owns agent loop timing, approval prompts, tool execution, session files, and how hook or MCP subprocesses are launched.
Connector role: agent-connector can only write the host's documented CLI config, prompt, memory, hook, MCP, or package files and point them at the home binary. Current wired connector surfaces: MCP server, Hooks, Skills, Memory.
Implementation consequence: A CLI surface is byte-addressable and smoke-testable from a temporary HOME, but terminal UI behavior still belongs to the host process. Visible host-native gaps for this surface: Commands, Subagents.

Verified Architecture

Source-checked on 2026-09-07. These notes are intentionally separate from the generated surface matrix below.

Sequence #40

Source pass

Current source review findings are kept as the first architecture section so readers can separate external evidence from the generated local coverage model.

Open Interpreter is a Rust fork of OpenAI's Codex; its config and hook subsystems are upstream Codex's under an isolated $INTERPRETER_HOME (default ~/.openinterpreter, $CODEX_HOME ignored) and a .openinterpreter project directory.
features.hooks is Stage::Stable and default_enabled, hooks live in hooks.json next to each config layer, and a live run against a mock provider fired SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, and SessionEnd with Claude-shaped stdin JSON.
Skills are documented at .agents/skills and ~/.agents/skills; subagents are [agents.<name>] config tables and no custom-prompts directory is documented, so those two surfaces remain unwired.

Runtime boundary

The adapter target is the local interpreter CLI. As a Codex fork it inherits Codex's agent loop, permission model, and config layout, but runs under its own home: the binary honors only $INTERPRETER_HOME, and the install script sets CODEX_COMMAND_NAME=interpreter and CODEX_HOME=$INTERPRETER_HOME.

The README states that Open Interpreter is a fork of OpenAI's Codex; the repository positions it for open models such as DeepSeek, Kimi, and Qwen.
The isolated home keeps an Open Interpreter install from reading or writing a Codex install on the same machine.
The row stays CLI-scoped rather than editor-extension scoped.

MCP registration

MCP is the one wired surface. The adapter writes [mcp_servers.<id>] tables into $INTERPRETER_HOME/config.toml: stdio servers carry command, args, and env; streamable-HTTP servers carry url with optional bearer_token_env_var and http_headers, the same shapes as codex-rs/config mcp_edit and mcp_types.

Local execution authority is the central risk boundary for any MCP tool exposed here.
TOML has no interpolation, so environment references resolve to literals at install time.

Hook bridge

Open Interpreter is a json-stdio hook host. As a Codex fork it ships Codex's hook subsystem with features.hooks on by default; the adapter writes the Claude-shaped hooks.json to ~/.openinterpreter (user) or <project>/.openinterpreter (trusted project), and a 2026-09-07 live run against a mock provider fired SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, and SessionEnd.

SessionEnd is registered on top of the codex event table because the interpreter product fires it (the host clamps its hook timeout to 3 seconds).
Non-managed command hooks run only after /hooks review in the TUI or with --dangerously-bypass-hook-trust, the same trust gate as upstream Codex.
Provider and model choice are host settings rather than hook semantics.

Content surfaces

Skills and memory are wired: skills go to the tool-neutral .agents/skills (project) and ~/.agents/skills (user), and memory is AGENTS.md under ~/.openinterpreter or the project root with AGENTS.override.md shadowing per directory. Subagents are [agents.<name>] tables in config.toml and no custom-prompts directory is documented, so those two stay host-native but unwired.

Open-model support makes provider configuration a first-class study axis.
Local code execution requires clear permission and sandbox language on the page.

Marketplace and host-only affordances

The host-only affordance is local openness: users can run with open models, control their environment, and execute code locally. That makes Open Interpreter useful for comparing terminal agents against cloud and editor hosts.

Model openness is a host affordance, not an agent-connector feature.
Computer-use behavior should remain separate from MCP and hook coverage.

Host-specific shape

Open Interpreter is a Rust Codex fork under its own $INTERPRETER_HOME: the adapter extends the codex adapter and wires MCP, Claude-shaped hooks (SessionEnd included), .agents skills, and AGENTS.md memory; subagents and custom prompts stay unwired.

Form factor: CLI.
Native surfaces: MCP server, Hooks, Commands, Skills, Subagents, Memory.
Wired surfaces: MCP server, Hooks, Skills, Memory.
Visible gaps: Commands, Subagents.

Hooks

Hooks enter through a host-launched command that exchanges event JSON over stdio.

MCP

The adapter writes this host's native MCP registration dialect, including its root key, scope, transport fields, and environment syntax.

Memory

Managed memory text is written into the rules or memory file this host actually reads, with ownership markers for reversible uninstall.

Marketplace and affordances

Current install is a direct native-file write path; any package artifact is manual unless a driver is added.

Runtime handlers: none.
User-visible affordances: Users see managed skills + memory surfaces after install; runtime-only surfaces stay behind the host boundary..

Known Limits

Local code execution authority needs careful safety wording.
Command and subagent paths stay unwired until the interpreter product documents them; hooks, skills, and memory were verified against the product on 2026-09-07.
Open model support is not equivalent to connector plugin support.

Evidence checked

Open Interpreter repository Open Interpreter docs

Architecture Diagram

Host-specific flow for the CLI surface.

json-stdio

1
Entry point: CLI
Terminal agent process reads user/project config and invokes the connector home binary for runtime callbacks.
2
Connector package
defineConnector package declares server, hooks, content, memory, and actions.
3
Adapter module: open-interpreter
json-stdio renderer at src/adapters/open-interpreter/index.ts.
4
Native host artifacts
Adapter writes MCP config + hook config + content files + memory/rules through this host's native file, package, or marketplace surface.
5
Runtime boundary
host launches JSON-stdio hook command
6
User-visible surface
Users see managed skills + memory surfaces after install; runtime-only surfaces stay behind the host boundary.
7
Coverage ceiling
4/6 native surfaces wired. Commands, Subagents are host-native but not wired yet.

Component Diagram

Static parts of the Open Interpreter integration and which side owns each one.

Open Interpreter host
CLI surface owns the model loop, UI shell, config discovery, and native lifecycle for this integration.
agent-connector package
Normalizes connector declarations into MCP, hook, memory, content, and action payloads before delegating to a host adapter.
open-interpreter adapter
Implements the json-stdio translation layer in src/adapters/open-interpreter/index.ts.
Native artifacts
Adapter writes MCP config + hook config + content files + memory/rules through this host's native file, package, or marketplace surface.
Runtime bridge
host launches JSON-stdio hook command
User-visible surface
Users see managed skills + memory surfaces after install; runtime-only surfaces stay behind the host boundary.

Runtime Sequence Flow

Message order from install to runtime for the CLI surface.

1 · User / project
Choose host adapter
Install or update starts against Open Interpreter; the selected surface is CLI.
2 · agent-connector
Resolve connector declaration
The package reads connector metadata and routes it to the open-interpreter adapter instead of exposing a generic host contract.
3 · open-interpreter adapter
Render host-native shape
Adapter writes MCP config + hook config + content files + memory/rules through this host's native file, package, or marketplace surface.
4 · Open Interpreter
Host loads artifacts
host launches JSON-stdio hook command
5 · Agent runtime
Run through host boundary
Users see managed skills + memory surfaces after install; runtime-only surfaces stay behind the host boundary.
6 · Coverage matrix
Expose remaining ceiling
4/6 native surfaces wired. Commands, Subagents are host-native but not wired yet.

Architecture Archive

The same five study axes are rendered for every host: runtime shape, hooks, MCP, memory, and host-only affordances.

Host-specific shape

Form factor: CLI.
Native surfaces: MCP server, Hooks, Commands, Skills, Subagents, Memory.
Wired surfaces: MCP server, Hooks, Skills, Memory.
Visible gaps: Commands, Subagents.

Hooks

Hooks enter through a host-launched command that exchanges event JSON over stdio.

MCP

The adapter writes this host's native MCP registration dialect, including its root key, scope, transport fields, and environment syntax.

Memory

Managed memory text is written into the rules or memory file this host actually reads, with ownership markers for reversible uninstall.

Marketplace and affordances

Current install is a direct native-file write path; any package artifact is manual unless a driver is added.

Runtime handlers: none.
User-visible affordances: Users see managed skills + memory surfaces after install; runtime-only surfaces stay behind the host boundary..

Special Feature Inventory

Host-specific capabilities separated from the generic surface matrix: lifecycle hooks, MCP registration, memory/rules, marketplace delivery, and UI or content affordances owned by the host.

Lifecycle hooks

Wired

Hooks enter through a host-launched command that exchanges event JSON over stdio.

MCP registration

Wired

The adapter writes this host's native MCP registration dialect, including its root key, scope, transport fields, and environment syntax.

Memory and rules

Wired

Managed memory text is written into the rules or memory file this host actually reads, with ownership markers for reversible uninstall.

Marketplace delivery

Direct files

Current install is a direct native-file write path; any package artifact is manual unless a driver is added.

Host-only affordances

1/3 wired

Wired host affordances: Skills. Host-native gaps: Commands, Subagents. Not confirmed: Statusline handler, Action affordances.

Source Review Notes

Latest source pass for host-specific architecture details. This is separate from generated adapter coverage and local drift guards.

Checked 2026-09-07Open review source

openinterpreter/open-interpreter repository (codex-rs/hooks, features, docs/hooks.md) and a live v0.0.41 run

Open Interpreter is a Rust fork of OpenAI's Codex; its config and hook subsystems are upstream Codex's under an isolated $INTERPRETER_HOME (default ~/.openinterpreter, $CODEX_HOME ignored) and a .openinterpreter project directory.
features.hooks is Stage::Stable and default_enabled, hooks live in hooks.json next to each config layer, and a live run against a mock provider fired SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, and SessionEnd with Claude-shaped stdin JSON.
Skills are documented at .agents/skills and ~/.agents/skills; subagents are [agents.<name>] config tables and no custom-prompts directory is documented, so those two surfaces remain unwired.

Evidence Map

Each host page ties public source evidence to the local adapter and drift guards that keep the rendered archive aligned with code.

Public source: openinterpreter/openinterpreter

Primary public repository or product source used for current host status and source links.

external

Open source

Local adapter implementation

Defines the install/uninstall behavior, native config rendering, capability flags, and any runtime parse/format bridge for this host.

local

src/adapters/open-interpreter/index.ts

Platform metadata and host-native matrix

Records the form factor, public source target, ranking source, native host surfaces, and current agent-connector coverage cells.

local

site/src/platform-data.ts

Architecture drift guard

Asserts that every registered host has a page, diagram, host-specific brief, source link, form-factor band, and adapter-aligned surface data.

local

tests/docs/platform-drift.test.ts

Runtime Shape

Hooks are dispatched through a host-launched command. The adapter parses host JSON, calls the connector handler, and formats the reply back to the host contract.

MCP registration: Wired through this adapter
Hook bridge: Wired

Connector Surfaces

Static content

MCP server, Hooks, Skills, Memory

Runtime handlers

No runtime handler surfaces wired

Surface Matrix

Wired means agent-connector installs it now. Host native marks support we know about but have not wired yet.

SurfaceStatus

MCP server

Static content

agent-connector installs this surface today.

Wired

Hooks

Static content

agent-connector installs this surface today.

Wired

Commands

Static content

The host appears to offer this, but this adapter has not wired it yet.

Host native

Skills

Static content

agent-connector installs this surface today.

Wired

Subagents

Static content

The host appears to offer this, but this adapter has not wired it yet.

Host native

Memory

Static content

agent-connector installs this surface today.

Wired

Statusline handler

Runtime handler

No confirmed host-native surface.

N/A

Action affordances

Runtime handler

No confirmed host-native surface.

N/A