This page collects agentic CLI tools that are useful when working with local coding agents, terminal sessions, and personal assistant runtimes.

It is not a replacement for the detailed guides. Use it as a map: pick the workflow you need first, then follow the specific post for setup commands, local trial notes, and caveats.

The Short Version

Tool Best For What It Is Not
OpenCode Provider-flexible open-source coding-agent CLI with local server/web/desktop surfaces A compose-first self-hosted dashboard
T3 Code Minimal web/desktop GUI for coding agents such as Codex, Claude Code, Cursor, and OpenCode A model provider or standalone agent runtime
Flue Building headless programmable agent harnesses in TypeScript A ready-made GUI or end-user coding agent
Codex CLI OpenAI-native local coding agent workflows A multi-provider neutral agent toolkit
Pi Running or building an agentic coding CLI with multi-provider LLM support A terminal multiplexer or compose-first web app
Herdr Managing persistent terminal workspaces for AI coding agents A model provider or a chat assistant runtime
NanoClaw Running a container-isolated personal assistant with CLI control A simple Docker Compose web app
OpenClaw Running a broader personal assistant Gateway with channels, providers, plugins, skills, and agents A narrow coding-only CLI
Hermes Agent Running a self-hostable AI agent runtime with memory, skills, gateway, cron, dashboard, and API server A small one-off prompt runner

OpenCode

OpenCode is an open-source AI coding agent CLI.

Use OpenCode when you want:

  • a provider-flexible coding agent with OpenAI, Anthropic, Google, OpenRouter, Azure, Bedrock, and other provider paths
  • terminal, local server, web, and desktop surfaces around the same tool
  • built-in build and plan agent modes
  • MCP, ACP, plugin, session, stats, GitHub, and model-listing commands
  • a Bun/TypeScript monorepo you can inspect and extend

The practical mental model: OpenCode is a full coding-agent product, not just a one-command wrapper. It can run in the terminal, expose a local server/web UI, and route through multiple model providers.

Useful links:

T3 Code

T3 Code is a minimal web and desktop GUI for coding agents.

Use T3 Code when you want:

  • a browser/desktop surface over existing coding-agent CLIs
  • Codex, Claude Code, Cursor, or OpenCode sessions from one interface
  • local server mode through npx t3@latest serve
  • terminal drawer and chat/project workflows
  • custom keybindings in ~/.t3/keybindings.json

The practical mental model: T3 Code is not the agent and not the model provider. It is the GUI/control surface you put on top of agents you have already installed and authenticated.

Useful links:

Flue

Flue is a TypeScript agent harness framework.

Use Flue when you want:

  • a programmable headless agent runtime rather than a GUI
  • workflows and agents defined in TypeScript
  • sessions, tools, skills, context, filesystem, and sandbox surfaces as framework primitives
  • Node or Cloudflare deployment targets
  • a way to build your own agent product rather than operate someone else’s CLI

The practical mental model: Flue is closer to Astro or Next.js for agents than to Codex CLI or OpenCode. You build with it; you do not mainly “chat in it.”

Useful links:

Codex CLI

Codex CLI is OpenAI’s local coding agent.

Use Codex CLI when you want:

  • the OpenAI-native terminal coding-agent path
  • interactive and non-interactive codex exec workflows
  • ChatGPT or API-key authentication
  • built-in sandbox modes such as read-only, workspace-write, and danger-full-access
  • MCP, plugin, app-server, remote-control, review, resume, and cloud-task surfaces
  • TypeScript or Python SDK integration around the Codex CLI

The practical mental model: Codex is the reference OpenAI coding-agent CLI. It is the agent you run; other tools can sit around it or feed it context.

Useful links:

Pi

Pi is an open-source agentic coding CLI and TypeScript toolkit.

Use Pi when you want:

  • an interactive terminal coding agent
  • print, JSON, or RPC modes for automation
  • OpenAI, Anthropic, Google, OpenRouter, Cloudflare, Bedrock, Copilot, and other provider paths
  • resumable sessions and session trees
  • skills, extensions, prompt templates, and themes
  • reusable packages for agent runtime, provider abstraction, and terminal UI work

The practical mental model: Pi can be the coding agent you run in the terminal, but it is also a toolkit for building agentic developer tools.

Useful links:

Herdr

Herdr is a terminal-native workspace manager for AI coding agents.

Use Herdr when you want:

  • persistent terminal panes for tools like Codex, Claude Code, or similar agent CLIs
  • a safer way to detach and reattach without killing the running session
  • status-aware panes for multiple agent sessions
  • remote attach and socket automation
  • keyboard shortcuts for repeated terminal-agent workflows

The practical mental model: Herdr is the control layer around your terminal agent sessions. It helps keep long-running CLI work organized.

Useful links:

NanoClaw

NanoClaw is a personal assistant runtime built around container-isolated agent execution.

Use NanoClaw when you want:

  • a local assistant runtime with separated admin and agent CLI surfaces
  • ncl for host-side administration and service control
  • claw as the terminal-facing assistant CLI after the NanoClaw skill is installed
  • container isolation for assistant workspaces
  • provider customization through NanoClaw skills such as Codex or OpenCode support

The practical mental model: NanoClaw is closer to an assistant runtime than a terminal multiplexer. It can expose CLI workflows, but it also manages agent execution, credentials, and containerized runtime behavior.

Useful links:

OpenClaw

OpenClaw is a local-first personal AI assistant Gateway.

Use OpenClaw when you want:

  • an always-on Gateway with CLI and dashboard control
  • chat/channel surfaces such as Telegram, Slack, Discord, Signal, Matrix, WhatsApp, and others
  • model-provider setup, failover, and runtime selection
  • agents, sessions, transcripts, plugins, skills, memory, and nodes
  • Docker or local daemon deployment options
  • diagnostic commands such as openclaw status, openclaw doctor, openclaw gateway health, and openclaw security audit

The practical mental model: OpenClaw overlaps with agentic CLI tools, but it is broader. It is a personal assistant control plane that can route messages, run agents, connect tools, and expose a local dashboard.

Useful links:

Hermes Agent

Hermes Agent is a self-hostable AI agent runtime from Nous Research.

Use Hermes Agent when you want:

  • a persistent assistant with memory, skills, scheduled jobs, and model routing
  • CLI, TUI, gateway, dashboard, and OpenAI-compatible API surfaces
  • Docker deployment with persistent state under /opt/data
  • messaging integrations and long-running assistant behavior
  • a broader runtime than a coding-only terminal agent

The practical mental model: Hermes Agent is not just a coding CLI. It is a personal agent runtime you can self-host, then connect to model providers and messaging surfaces. The local pass validated Docker Compose and source structure, but not a complete live model-backed conversation.

Useful links:

Choosing One

If You Need… Start With
A provider-flexible open-source coding agent OpenCode
A local web/server surface for an agentic CLI OpenCode
A minimal GUI over existing coding-agent CLIs T3 Code
A TypeScript framework for building headless agents Flue
The OpenAI-native local coding agent Codex CLI
Codex SDK or app-server integration Codex CLI
A coding-agent CLI with multi-provider model support Pi
Reusable TypeScript packages for agent runtime and TUI work Pi
Persistent panes for multiple coding agents Herdr
A clean detach and reattach flow for terminal sessions Herdr
A personal assistant runtime with containerized execution NanoClaw
A host admin CLI for an assistant service NanoClaw ncl
A terminal assistant CLI installed into the runtime NanoClaw claw
A broader personal assistant Gateway with channels and a dashboard OpenClaw
A self-hosted agent runtime with memory, skills, cron, gateway, dashboard, and API server Hermes Agent
A simple compose-first self-hosted web UI Neither is the main fit

Local Validation Notes

Tool What Was Validated Locally Remaining Caveat
OpenCode installed CLI help, agent command help, OpenAI model catalog listing, and local server/web shell on port 14096 source dev was blocked because Bun was not installed; no authenticated model call and no Docker Compose deployment were validated
T3 Code npx t3@latest --help, published package metadata, local server startup, migrations, listening URL, and pairing URL no completed browser session and no authenticated Codex/Claude/Cursor/OpenCode run
Flue package metadata, CLI help output, and flue init --target node scaffold help returned exit 1; no live model-backed workflow or workspace build due resource limits
Codex CLI installed CLI help, no-auth codex exec failure mode, workspace dependency install, TypeScript SDK build no authenticated Codex run and no Rust source build on this machine
Pi npm install, monorepo build, source-run CLI help, no-key failure path, and upstream no-key test script No authenticated provider session was run, and Pi is not sandboxed by default
Herdr Local terminal workflow, key shortcuts, detach behavior, and practical usage notes It is not a Docker Compose service, so there is no Home-Lab compose file to publish
NanoClaw Host install, typecheck, test suite, and ncl help path Full installer, container runtime, credentials, and a live assistant conversation were not completed locally
OpenClaw published CLI help, gateway help, no-setup status/doctor diagnostics, upstream compose config, and Home-Lab compose config no live authenticated model call; no image pull/build/start due disk/RAM pressure
Hermes Agent source inspection, upstream compose rendering, Home-Lab compose rendering, pinned image tag, and API exposure caveats no live authenticated model call; local CLI smoke test was blocked by a stale launcher and missing ambient Python dependencies

OpenAI And Provider Notes

OpenCode can list OpenAI models and its provider/model spec includes openai, but the local pass did not run an authenticated OpenAI task.

Codex CLI is the OpenAI-native path. The local post validated the CLI surface and an empty-auth failure mode, but did not spend tokens on an authenticated task.

Pi supports OpenAI API-key usage via OPENAI_API_KEY and also documents ChatGPT Plus/Pro Codex login through /login. During the local pass, I validated the source build, CLI help, and no-key failure mode, but not a live OpenAI session.

NanoClaw can be extended toward OpenAI-compatible workflows through its Codex or OpenCode provider skills, but current upstream behavior should be treated carefully:

  • the default local checkout was validated around Claude-oriented defaults
  • Codex support is skill-driven and requires extra setup, credentials, and image rebuilds
  • OpenCode support can route through providers such as OpenAI-compatible endpoints, depending on configuration
  • we did not validate a live OpenAI API run locally during the first pass

That distinction matters for posts and docs: provider support is different from having the provider ready out of the box.

OpenClaw documents OpenAI provider support and a Codex app-server runtime path through openai/* model refs. The local pass validated CLI and config behavior, not an authenticated OpenAI or Codex turn.

T3 Code depends on the underlying coding-agent CLI. For Codex use, install/authenticate Codex first with codex login, then use T3 Code as the GUI surface.

Hermes Agent can expose an OpenAI-compatible API server, but the local pass only validated Compose configuration. Keep API_SERVER_ENABLED=false until you configure a strong API_SERVER_KEY and a trusted network boundary.