Terminal AI Agents: The 2025 Landscape
Table of Contents
- 1. Background
- 2. Industry Context
- 3. GitHub Stars Leaderboard (December 2025)
- 4. The Big Three: Enterprise Terminal Agents
- 5. Open Source Champions
- 6. Ecosystem Map
- 7. Emerging Protocols
- 8. Benchmarks
- 9. FreeBSD Compatibility Testing
- 10. Methodology
- 11. Repository
- 12. Conclusion
- 13. References
- 14. Related notes
- 15. Postscript (2026)
1. Background
Since February 2025, Anthropic, Google, and OpenAI have all released command-line coding tools.
This research documents hands-on evaluation of 20+ terminal AI agents, including installation testing on FreeBSD 14.3, GitHub stars tracking, and feature comparison.
"Our big bet is that there's a future in which 95% of LLM-computer interaction is through a terminal-like interface." — Mike Merrill, Co-creator of Terminal-Bench (Stanford)
2. Industry Context
a16z 53% and Gartner 75% claims lack direct source links; statistics appear plausible but require verification
2.1. Adoption Statistics
Direct fetch of Gartner's own 2024-04-11 press release ("Gartner Says 75% of Enterprise Software Engineers Will Use AI Code Assistants by 2028") returned HTTP 403, but a search-engine summary of that same page corroborates its own title and text -- the 75%-by-2028 and <10%-in-2023 figures as cited. The a16z figure does not hold up the same way: a direct fetch of a16z.com/generative-ai-enterprise-2024/, the report named inline, found no 53% experimentation-to-production statistic anywhere in that document -- the cited number is not traceable to the named source.
- 53% of companies experimenting with LLMs for development have deployed to production (a16z Enterprise AI Report 2024)
- 75% of enterprise software engineers will use AI assistants by 2028, up from <10% in 2023 (Gartner 2025)
3. GitHub Stars Leaderboard (December 2025)
Live GitHub API re-check (2026-10-08, api.github.com/repos/<org>/<repo>) against the December-2025 snapshot above: Claude Code 48,023 -> 149,701 stars, Gemini CLI 88,306 -> 107,255, Aider 39,129 -> 49,427, Efrit 338 -> 455. The ranking has flipped since the snapshot -- Claude Code now leads Gemini CLI outright, reversing the #7/#1 order recorded here -- confirming the table's own framing as a dated snapshot rather than a current ranking; no number in the table itself was found wrong for its stated date.
Rankings based on live GitHub API data collected during this research:
| Rank | Agent | Stars | Provider |
|---|---|---|---|
| 1 | Gemini CLI | 88,306 | |
| 2 | OpenHands | 65,846 | Open Source |
| 3 | Open Interpreter | 61,209 | Open Source |
| 4 | Cline | 56,318 | Open Source |
| 5 | GPT Engineer | 55,131 | Open Source |
| 6 | OpenAI Codex CLI | 54,514 | OpenAI |
| 7 | Claude Code | 48,023 | Anthropic |
| 8 | Aider | 39,129 | Open Source |
| 9 | Continue | 30,453 | Open Source |
| 10 | Goose | 24,922 | Block/Square |
| 11 | Devika | 19,477 | Open Source |
| 12 | SWE-agent | 18,075 | Princeton NLP |
| 13 | Plandex | 14,781 | Open Source |
| 14 | gptme | 4,089 | Open Source |
| 15 | Efrit | 338 | Steve Yegge |
// Terminal AI Agents 2025 — framework comparison matrix
// Compares the most-cited agents from the survey across provider type,
// install ecosystem, key differentiator, and FreeBSD 14.3 compatibility.
// Build: dot -Tpng diagram-comparison.dot -o diagram-comparison.png
digraph terminal_agents_comparison {
rankdir=TB;
graph [bgcolor="white", fontname="Helvetica", fontsize=11,
pad="0.3", nodesep="0.3", ranksep="0.4", compound=true];
node [shape=box, style="rounded,filled", fontname="Helvetica",
fontsize=10];
edge [fontname="Helvetica", fontsize=9, color="#888888"];
// Column legend (every row below uses this order)
legend [label="{ Agent | Stars (Dec 2025) | Install | Key differentiator | FreeBSD 14.3 }",
shape=record, fillcolor="#fafafa", color="#444444", fontcolor="#444444"];
// Enterprise tier — "The Big Three"
subgraph cluster_enterprise {
label="Enterprise tier (vendor-backed)"; labeljust="l";
style="rounded,filled"; color="#1d4ed8";
fontcolor="#1d4ed8"; bgcolor="#eff6ff";
cc [label="{ Claude Code | 48,023 | npm i -g @anthropic-ai/claude-code | MCP, 5 permission modes, SWE-bench 72%+ | works }",
shape=record, fillcolor="#dbeafe", color="#1d4ed8", fontcolor="#1d4ed8"];
gem [label="{ Gemini CLI | 88,306 | npm i -g @google/gemini-cli | 1k req/day free, 1M-token context | works }",
shape=record, fillcolor="#dbeafe", color="#1d4ed8", fontcolor="#1d4ed8"];
cdx [label="{ OpenAI Codex CLI | 54,514 | npm i -g @openai/codex | ChatGPT subscription, lightweight | fails (explicit FreeBSD check) }",
shape=record, fillcolor="#dbeafe", color="#1d4ed8", fontcolor="#1d4ed8"];
}
// Open-source champions
subgraph cluster_oss {
label="Open-source champions"; labeljust="l";
style="rounded,filled"; color="#15803d";
fontcolor="#15803d"; bgcolor="#f0fdf4";
oh [label="{ OpenHands | 65,846 | (containerized) | end-to-end software-engineering agent | n/a (container) }",
shape=record, fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
oi [label="{ Open Interpreter | 61,209 | uv tool install open-interpreter | exec() loop for Python/JS/shell | works (44 min build) }",
shape=record, fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
cln [label="{ Cline | 56,318 | (VS Code extension) | autonomous coding in editor | n/a }",
shape=record, fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
aid [label="{ Aider | 39,129 | pip install aider-chat | git-native, 84.9% polyglot accuracy | fails (tree-sitter parser.h) }",
shape=record, fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
cnt [label="{ Continue | 30,453 | npm i -g @continue/cli | TUI + headless, custom assistants | not tested }",
shape=record, fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
}
// Specialist / niche tier
subgraph cluster_niche {
label="Specialist tier"; labeljust="l";
style="rounded,filled"; color="#6b21a8";
fontcolor="#6b21a8"; bgcolor="#f5f3ff";
gse [label="{ Goose | 24,922 | (Block/Square) | enterprise workflows | issues (langfuse.decorators) }",
shape=record, fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
amp [label="{ Amp | n/a | binary from ampcode.com | Sourcegraph agent, Go binary | works }",
shape=record, fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
ghc [label="{ GitHub Copilot CLI | n/a | gh extension install github/gh-copilot | command suggest + explain via gh | works }",
shape=record, fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
gpt [label="{ gptme | 4,089 | pip install gptme | shell + Python + browser + vision | not tested }",
shape=record, fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
efr [label="{ Efrit | 338 | (Emacs package) | pure Elisp, Steve Yegge | works }",
shape=record, fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
}
// Cross-cutting protocol callout
proto [label="Emerging protocols\nMCP (Anthropic) — Claude Code, Cursor, Cline, Continue\nACP (Zed) — Apache-licensed editor/agent standard",
shape="note", fillcolor="#fef9c3", color="#a16207",
fontcolor="#a16207"];
// Invisible edges keep legend → enterprise → oss → niche → protocols stacked
legend -> cc [style=invis, lhead=cluster_enterprise];
cc -> oh [style=invis, lhead=cluster_oss, ltail=cluster_enterprise];
oh -> gse [style=invis, lhead=cluster_niche, ltail=cluster_oss];
gse -> proto [style=invis, ltail=cluster_niche];
}
4. The Big Three: Enterprise Terminal Agents
Direct fetch of anthropic.com/news/claude-4 confirms the figure: "Claude Opus 4 is our most powerful model yet and the best coding model in the world, leading on SWE-bench (72.5%)," announced 2025-05-22 (without extended thinking). That resolves the prior needs-citation gap -- Claude Code's "72%+" is this Opus 4 SWE-bench Verified score, not a Claude-Code-specific benchmark run. Figure was current as of this note's 2025-12-22 date; the model family has since moved past it (see Postscript).
4.1. Claude Code (Anthropic)
- Released: February 24, 2025
- Stars: 48,023
- SWE-bench: 72%+ (leading published scores)
- Install:
npm install -g @anthropic-ai/claude-code
Key differentiators:
- Multi-tool orchestration with MCP protocol
- 5 permission modes for enterprise control
- JSON schema validation for structured output
- Docker sandboxing and offline capability
4.2. Gemini CLI (Google)
- Stars: 88,306 (highest among terminal agents)
- Free Tier: 1,000 requests/day with 1M-token context window
- Install:
npm install -g @google/gemini-cli
Key differentiators:
- Most generous free tier
- Native Google service integration
- Sandbox execution with custom containers
4.3. OpenAI Codex CLI
- Released: April 16, 2025
- Stars: 54,514
- Install:
npm install -g @openai/codex
Key differentiators:
- Works with ChatGPT subscription (cost-effective)
- Lightweight and fast
- Full customizability
5. Open Source Champions
5.1. Aider - The Git-Native Agent
- Stars: 39,129
- Accuracy: 84.9% on polyglot test suite
- Install:
pip install aider-chat
Aider's repository mapping feature understands entire codebases. It automatically commits changes with sensible messages and supports voice input.
5.2. Open Interpreter - Natural Language Computing
- Stars: 61,209
- Version: 0.4.3
- Install:
uv tool install open-interpreter
Equips LLMs with an exec() function for running code in Python, JavaScript, and shell. Streams output to terminal.
5.3. Continue - The Extensible Platform
- Stars: 30,453
- Enterprise Users: Siemens, Morningstar
- Install:
npm install -g @continue/cli
TUI mode for interactive sessions, headless mode for automation. Custom assistant framework with community hub.
5.4. gptme - Personal AI Agent
- Stars: 4,089
- Install:
pip install gptme
Shell, Python, browser, and vision tools. Self-correction feedback loop. Works as GitHub bot or web interface.
5.5. GitHub Copilot CLI
- Docs: GitHub Copilot Documentation
- Install:
gh extension install github/gh-copilot
GitHub's official CLI extension for Copilot. Integrates with gh CLI for command suggestions and explanations. Requires GitHub Copilot subscription.
5.6. Amp - Sourcegraph's Terminal Agent
- Docs: Amp CLI Manual
- Install: Download from ampcode.com
Sourcegraph's agentic coding assistant. Runs as CLI or integrates with VS Code. Go binary works natively on FreeBSD.
6. Ecosystem Map
// Terminal AI agents 2025 -- vendor clusters with shared protocol bridges
digraph terminal_ai_ecosystem {
rankdir=LR;
graph [bgcolor="white", fontname="Helvetica", fontsize=11,
pad="0.3", nodesep="0.3", ranksep="0.55"];
node [shape=box, style="rounded,filled", fontname="Helvetica",
fontsize=10, fillcolor="#dbeafe", color="#888", fontcolor="#555"];
edge [color="#888", fontcolor="#555"];
subgraph cluster_anthropic {
label="Anthropic"; style="rounded"; color="#b91c1c"; fontcolor="#b91c1c";
claude_code [label="Claude Code", fillcolor="#fee2e2", color="#b91c1c", fontcolor="#b91c1c"];
}
subgraph cluster_openai {
label="OpenAI"; style="rounded"; color="#b45309"; fontcolor="#b45309";
codex_cli [label="Codex CLI", fillcolor="#fef3c7", color="#b45309", fontcolor="#b45309"];
}
subgraph cluster_google {
label="Google"; style="rounded"; color="#1d4ed8"; fontcolor="#1d4ed8";
gemini_cli [label="Gemini CLI", fillcolor="#dbeafe", color="#1d4ed8", fontcolor="#1d4ed8"];
}
subgraph cluster_oss {
label="Open Source"; style="rounded"; color="#15803d"; fontcolor="#15803d";
aider [label="Aider", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
cline [label="Cline", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
continue [label="Continue", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
openhands [label="OpenHands", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
oi [label="Open Interpreter", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
goose [label="Goose (Block)", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
gptme [label="gptme", fillcolor="#dcfce7", color="#15803d", fontcolor="#15803d"];
}
subgraph cluster_protocols {
label="Shared Protocols"; style="rounded"; color="#6b21a8"; fontcolor="#6b21a8";
mcp [label="MCP\n(Model Context Protocol)", fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
acp [label="ACP\n(Agent Client Protocol)", fillcolor="#ede9fe", color="#6b21a8", fontcolor="#6b21a8"];
}
claude_code -> mcp;
cline -> mcp;
continue -> mcp;
codex_cli -> mcp;
goose -> mcp;
gemini_cli -> mcp;
continue -> acp;
cline -> acp;
aider -> acp [style=dashed];
}
7. Emerging Protocols
7.1. MCP (Model Context Protocol)
Anthropic's protocol for tool integration, now adopted by:
- Claude Code
- Cursor CLI
- Cline
- Continue
7.2. ACP (Agent Client Protocol)
Zed's open standard for agent-editor integration:
- Multi-IDE support (Zed, JetBrains coming)
- No vendor lock-in
- Apache licensed
8. Benchmarks
benchmark definitions are accurate reference material; SWE-bench Pro performance gap (23% vs 72%) correctly reflects real-world complexity jump
| Benchmark | Focus | Top Performance |
|---|---|---|
| SWE-bench | GitHub issue resolution | Claude Code 72%+ |
| SWE-bench Pro | Harder variant | GPT-5/Claude Opus ~23% |
| Terminal-Bench | Real terminal tasks | Featured on Claude 4 card |
| LiveBench | Continuously updated | Varies by model |
9. FreeBSD Compatibility Testing
Hands-on testing on FreeBSD 14.3-RELEASE:
| Agent | Status | Notes |
|---|---|---|
| Claude Code | ✅ Works | Native Node.js support |
| Amp | ✅ Works | Go binary |
| Gemini CLI | ✅ Works | Native npm |
| GitHub Copilot CLI | ✅ Works | Via gh extension |
| Efrit | ✅ Works | Pure Elisp |
| Open Interpreter | ✅ Works | v0.4.3, 44min build |
| Aider | ❌ Fails | tree-sitter parser.h missing |
| OpenAI Codex CLI | ❌ Fails | Explicit FreeBSD check |
| Goose | ⚠️ Issues | langfuse.decorators missing |
10. Methodology
This research was conducted using Claude Code itself as the primary interface. The approach:
- Discovery: Web search for terminal AI agents released in 2024-2025
- Star Collection: GitHub API queries for live star counts
- Installation Testing:
uv tool installandnpm install -gon FreeBSD - Feature Extraction:
--helpoutput analysis for each agent - Documentation: Org-mode literate programming
All data collected December 22, 2025.
11. Repository
Full comparison matrices, architecture diagrams, and workshop labs available at:
12. Conclusion
Claude Code leads in reasoning and enterprise features. Gemini CLI wins on free tier generosity. Aider dominates git-native workflows. The open-source ecosystem (OpenHands, Cline, Continue) offers unmatched customization.
13. References
15. Postscript (2026)
The 2025 snapshot held up surprisingly well into Q2 2026, but several plates have shifted. Anthropic shipped the Claude 4.x family with extended-thinking modes that pushed Claude Code's SWE-bench numbers further ahead, and reframed sub-agents as first-class primitives via Code Mode. OpenAI surprised the industry by releasing the gpt-oss open-weight family, narrowing the "open vs closed" gap that defined the 2025 leaderboard and giving Open Interpreter, Aider, and gptme credible local backends. MCP has crossed into enterprise default-on territory – every agent in the table above now ships an MCP client, and most cloud vendors expose first-party MCP servers (AWS, Cloudflare, GitHub). The terminal-vs-IDE framing from 2025 has largely collapsed: Cursor, Zed, and VS Code all embed terminal agents, while Claude Code and Codex CLI ship IDE-aware modes.