Things on this page are fragmentary and immature notes/thoughts of the author. Please read with your own judgement!
AI Agent Code CLI¶
Most opensource AI agent code cli tools
are LLM agnostic
support multi-model mode
Table 1:AI Agent Code CLIs Comparison
Agent CLI | LLM Agnostic? | Multi-Model? | Remote API? | Mobile App? | Primary Focus | Language | License | Popularity (Stars) |
|---|---|---|---|---|---|---|---|---|
Y | Y | Y[1] | Y[2] | Full ecosystem (CLI, Desktop, Mobile, SDK) | TypeScript | MIT | ~192.7k ★ | |
Y | Y[3] | Y[4] | N | Minimalist, extensible harness | TypeScript | MIT | ~82.8k ★ | |
Y | Y | Y[5] | N[6] | General OS execution / testing | Rust / TS | Apache 2.0 | ~52.2k ★ | |
Y[7] | Y[8] | Y[9] | N | Git-integrated pair programming | Python | Apache 2.0 | ~47.3k ★ | |
N[10] | Y[11] | Y[12] | N[13] | DeepSeek prefix-cache optimization | Go / TS | MIT | ~31k ★ | |
Y[14] | Y[15] | N[16] | N | Elegant TUI with LSP integration | Go | MIT | ~27k ★ | |
Y | Y[17] | Y[18] | N | Multi-step planning for large tasks | Go | MIT | ~15.6k ★ | |
Y | Y[19] | Unknown | N | Hash-anchored edits, subagents | Python / TS | Open Source | ~11.6k ★ | |
Y[20] | Y[21] | Unknown | N | Multi-model MCP-driven coding | Rust | Apache 2.0 | ~7.5k ★ | |
Y | Y[22] | Y[23] | N | High-perf, zero-unsafe isolation | Rust | Open Source | ~1.5k ★ | |
Codex CLI (OpenAI Official) | N[24] | N[25] | Y[26] | N | Enterprise proxy/sandbox workflows | Rust | Apache 2.0 | N/A |
N[27] | Y[28] | Y[29] | Y[30] | Native terminal integration w/ MCP | TypeScript | Closed | N/A |
References¶
JS/TS SDK.
Dedicated OpenCode Mobile app.
toolModel: separate reasoning vs. tool-calling model.RPC / SDK.
REST API.
Desktop & CLI only, no mobile app.
100+ providers via LiteLLM.
--model/--editor-modelflags.--messageflag for headless/one-shot runs.DeepSeek models only.
Executor + planner via
planner_model.JSON-RPC interface.
Desktop / VS Code integration only, no mobile app.
Multi-model support across providers.
Large model for the Coder agent, small model for the Task agent.
Focuses on the TUI; no dedicated remote API.
Per-role model packs: planner/architect/coder/builder/summarizer.
Client/server architecture.
Role-based routing, e.g. a dedicated
planrole.300+ models via OpenRouter and direct providers.
Forge/Muse/Sage agents, each configurable with its own model.
Ports Pi’s
toolModeldual-model design.Built-in HTTP server.
Locked to OpenAI models officially; community forks add other providers.
Plan mode inherits the global model; manual switching only, no built-in per-role split.
Python SDK.
Anthropic models only (official).
Per-subagent model configuration.
Via MCP servers and hooks.
Session handoff to mobile via QR code.