2026 年中時点の CLI コーディングエージェントの現状（37 分読了）

The terminal was an unlikely winner. In 2024 the bet was on IDEs — Copilot already lived in the editor, Cursor was climbing, and "agent" still meant chat with shell access bolted on. By mid-2026 the serious usage ran from CI, SSH, and machines with no GUI at all. Scriptable, boring, and the model never fights something else for the window manager.

35 actively maintained CLI coding agents as of July 3, 2026. Crowded field.

Skip to feature by feature comparison · Skip to conclusions

Short history

The 1st wave arrived in 2023, before anyone called these things agents. gptme (March 2023) let a model run shell commands from the terminal. Aider (mid-2023) built AI pair programming around git, with atomic commits as the unit of change. Open Interpreter (July 2023) generalized the idea to controlling the whole computer. All 3 survive — gptme as a daemon, Aider as a pair programmer, Open Interpreter as a general computer controller.

Anthropic's Claude Code research preview in February 2025 set the shape: agentic loop, file and shell tools, project memory file, permission prompts, plan mode, hooks, subagents. Everyone else cloned it. OpenAI shipped Codex CLI in April 2025 (later rewritten in Rust). Google followed with Gemini CLI in June 2025 and an aggressive free tier. By the 2nd half of 2025 the releases piled up in one quarter — Cursor, Amp, Augment, Factory, Charm, and 12 open-source teams.

Teams that standardized on Gemini CLI in 2025 still have configs from last fall. December 2025 is when the standards people showed up: the Linux Foundation formed the Agentic AI Foundation (AAIF), anchored by Anthropic's Model Context Protocol, OpenAI's AGENTS.md convention, and Block's Goose agent, with backing from Google, Microsoft, AWS, Cloudflare, and Bloomberg. The same month, Sourcegraph spun its agent out as the independent Amp Inc., Mistral entered with Vibe and the Devstral 2 models, and Codebuff open-sourced its multi-agent harness.

The 1st half of 2026 was messier. Copilot CLI hit GA in February. Cline shipped CLI 2.0 and peeled off an open SDK. Kilo CLI reached 1.0. Junie went from beta to GA in June. Google stole the headline at I/O in May: Antigravity CLI launched, and Gemini CLI got a kill date — June 18 for free and consumer-plan users. Grok Build landed late May. Moonshot replaced its earlier CLI with Kimi Code in June.

Models kept pace. DeepSeek V4 under MIT in April. GLM-5.2 and Kimi K2.7 Code in June, pulling open weights to within a few points of frontier terminal scores. The free tiers started dying around the same time: Qwen Code's hosted tier in April, Amp Free waitlisted, Gemini's consumer path gone by summer.

Model labs

Lab agents come from the model vendors — harness and model in one box, usually on a subscription line item already there.

Claude Code still sets the vocabulary — agent teams, hooks, skills, the lot. Experimental agent teams (multiple sessions messaging each other) remain ahead of the pack. The trade is total model lock-in. As of June 2026, headless and SDK usage bill from a separate credit pool. Terminal-Bench 2.0 in early July had Anthropic's newest models on top. The bundled pitch is a co-trained model plus harness, and the leaderboard backs it.

Codex CLI is the strongest counterweight, and it had a busy spring. It's open source, OS-sandboxed, rides ChatGPT's subscriber base, and runs GPT-5.5 as its current frontier default. Through the 1st half of 2026 it picked up persistent Goals with token budgets, thread-level delegation to subagents, a plugin marketplace, browser use, encrypted remote execution, and a one-liner to import Claude Code configuration — switching costs are real enough that Codex had to automate them.

If you standardized on Gemini CLI in 2025, June 18 was a bad day. Google grew it into one of the most-starred repos in the category, then pointed free and consumer-plan users at the closed-source Antigravity CLI instead. Antigravity is a Go rewrite that shares its harness with the Antigravity desktop platform (Gemini 3.5 at the core), runs async multi-agent workflows, and is free in public preview, with non-Google models on the menu. Free tiers are not foundations.

Grok Build, Kimi Code, and Mistral Vibe showed up late — plan mode, subagents, headless CI already checked off. Grok Build shipped in beta with plan mode, worktree-isolated parallel subagents, and headless CI support on day 1. Kimi Code and Mistral Vibe both lean on unusually cheap, partly open models: Mistral's Devstral 2 posts a vendor-reported 72.2 percent on SWE-bench Verified from a 123B model, and Moonshot's K2.7 Code, released mid-June with open weights, undercuts frontier pricing by an order of magnitude. Both adopted ACP so any compatible editor can host them. Qwen Code is the fork case: a Gemini CLI derivative retuned for Qwen's open-weight coders, still alive after its free tier died because the tool is Apache-licensed and endpoint-agnostic.

Platform and product CLIs

Platform vendors sell the agent as part of a development stack — distribution, integration, governance. Almost all of them are multi-model.

Copilot CLI wins on reach. GA in February 2026, $10/mo entry tier, GitHub's MCP server built in, plugins from repos, per-repo memory, specialist agents for explore/plan/review/build. Prefixing a prompt with & pushes the job to GitHub's cloud coding agent. For GitHub-centric teams it's the lazy default, and its model picker makes it one of the most vendor-neutral closed agents. On July 1 Copilot added Moonshot's open-weight Kimi K2.7 Code on Azure — the first open-weight model inside a major closed platform agent.

After Copilot, the bets split.

Cursor CLI is the consistency play — same agent, same rules, same MCP config in the IDE, the terminal, and CI. One AGENTS.md driving all three is the whole product.

Amp drops the model picker entirely and swaps models when its team decides the mix should change. The ad-funded Amp Free tier (sponsor-backed, no-training guarantee) is the strangest business model in the list; admission is closed for now.

Auggie front-loads context: index the whole repo before the first prompt. Pays off on legacy monoliths, less on greenfield repos.

Droid is the enterprise bundle — specialist agents, heavy parallelism, Slack and ticketing hooks. Factory posted the best Terminal-Bench numbers of any product harness in late 2025.

Junie drags JetBrains' debugger and database wiring to the terminal over ACP.

Qoder and CodeBuddy are the China stack plays — same 2026 feature checklist as everyone else, different distribution behind the firewall.

Open-source harnesses

Bring-your-own-key agents are the largest group, and open-weight pricing changed what they cost to run. A decent harness plus GLM, DeepSeek, Qwen3-Coder, Devstral, or Kimi tokens gets you frontier capability for a fraction of the subscription price. The gap closed sharply in June, when Z.ai released GLM-5.2 under MIT: a 744B mixture-of-experts model (about 40B active) with a 1 million-token context that beats GPT-5.5 on several long-horizon coding benchmarks at roughly 1/6 of the price (Terminal-Bench 2.1 analysis). It also runs fully offline on high-memory hardware, needing around 245GB of memory at the most aggressive usable quantization.

OpenCode has the most GitHub stars in the category at 182k — ahead of Gemini CLI and Codex — with 75+ providers, a client-server architecture, and a polished TUI plus desktop and IDE surfaces. It also has the messiest fork story: the original author joined Charm in mid-2025, where that lineage became Crush (prettier TUI, source-available FSL license), while the SST team's OpenCode kept growing. Two healthy projects out of one messy split.

DeepSeek-Reasonix is the fastest riser: 26k stars in roughly 10 weeks from one static Go binary. The design treats inference cost as an engineering problem. Sessions keep DeepSeek's byte-stable prefix cache hot — stable env summaries at startup, stale tool output pruned before compaction, planner and executor on separate cache-stable threads.

One reported day: 435M input tokens, 99.82% cache hits, about 12.Uncachedestimate:12. Uncached estimate: 61. Any OpenAI-compatible endpoint, MCP over stdio and HTTP, checkpoints, rewind, and chat-ops bridges to Feishu, Lark, and WeChat. DeepSeek V4 landing the same April under MIT did not hurt.

Goose took a different route to durability: Block donated it to the Agentic AI Foundation, making it the only major agent under neutral foundation governance, alongside the MCP and AGENTS.md standards themselves. Cline matters beyond its user base because of its Apache-licensed SDK, extracted in May 2026, which turned a popular product into infrastructure others can build on; GitHub (Copilot SDK), Anthropic (Agent SDK), and OpenHands (a research-grade Python SDK with deterministic replay) made the same move. Kilo and Continue extend IDE-extension franchises into the terminal — Kilo with orchestration modes and 500+ models, Continue with a CI-first angle — though Continue's 2026 commit cadence puts it on the watch list.

Aider invented half the git discipline everyone else copied, and its polyglot benchmark shaped model evaluation for 2 years. It stayed a pair-programming tool while the market went agentic, though, and releases have slowed to a trickle. Still strong at atomic commits; the category moved elsewhere.

The long tail is larger than the headline tools suggest. Every Code out-features its own upstream, orchestrating OpenAI, Anthropic, and Google models on a Codex base. Codebuff decomposes work into explicit finder, planner, editor, and reviewer agents, and claims 61 percent versus 53 for Claude Code on its own 175-task evaluation. Kode lets a session consult a specific model by name mid-task. Nanocoder is for people who want collectively owned, telemetry-free, local-first — no company behind it.

Minimal cores and pioneers

Pi is the backlash. Mario Zechner argues a modern model needs 4 tools (read, write, edit, bash), a system prompt under 1,000 tokens, and nothing else in core — MCP included belongs in a TypeScript extension SDK. Personal project to 67k stars in under a year, which says the audience for harness bloat had thinned.

oh-my-pi — Omp for short — is the opposite bet: a batteries-included Pi distribution that kept adding features until it had the widest feature set in the comparison below. Same community, opposite directions. Pi bets on subtraction, Omp bets on accumulation, and both arguments still have an audience.

mini-SWE-agent is the control group: about 100 lines of Python, bash only, and strong frontier models still clear a large share of SWE-bench. Hard to justify a 50-feature harness without beating that. gptme and Open Interpreter round out the elders — gptme still in active development and focused on persistent, self-improving agents; Open Interpreter historically important and visibly slowing.

Feature by feature

By 2025 every serious agent had the same skeleton: edit-run-test loop, project instructions file, MCP, plan mode, permission prompts, headless mode. None of that separates anyone anymore.

What marketing pages still hide is the rest. Five harnesses get the full treatment — Claude Code, Codex CLI, OpenCode, Omp, plus Copilot CLI as the deployed-at-scale platform pick. Everyone else appears in the notes under each table.

Context and memory

A year ago project memory meant a markdown file someone on the team updated by hand. Now 3 of the 5 leaders learn on their own: Claude Code accumulates memory files across sessions, Copilot builds a per-repository understanding, and Omp writes facts mid-run and synthesizes them later on request. Omp also reads the rules, skills, and MCP registrations other agents leave in a repo. Switching over barely touches config. Elsewhere, Kilo's Memory Bank stores agent state in structured markdown inside the repo, and gptme goes furthest of all with git-backed memory for agents that are meant to run for months.

Editing and code intelligence

Omp leads here by a lot. Hash-anchored editing is aimed at the refactor failure everyone hits — the patch lands on the wrong line because whitespace moved. AST rewrites with preview-then-accept, plus a debugger the agent can drive. No lab agent ships that as of July 2026. Heavier setup than Claude Code; you still pick the model. This is where to look when patches keep landing on the wrong line. The rest of these tools approach code intelligence differently: Auggie's context engine builds a semantic index before work begins, Aider's repo map compresses project structure into the prompt, ForgeCode does embedding-based search after a sync step, and Junie leans on the JetBrains indexer — probably the deepest static analysis any agent gets, just not an open one.

Orchestration

Claude Code and Omp let agents coordinate as peers — agent teams on one side, an in-process chat channel plus a standing Advisor model on the other. Codex and Copilot push parallelism to the cloud — less load on the laptop, more on the ops queue. Grok Build isn't in the table but shipped worktree-isolated parallel subagents on day 1 (up to 8). Codebuff hardcodes roles. Mistral Vibe keeps remote agents running after the terminal closes. Amp's oracle — a stronger model for hard steps — is the nearest cousin to Omp's Advisor. Scheduling showed up late: recurring cloud agents on Claude Code, timed reminders on Codex in June, CI everywhere else.

Extensibility

MCP for external tools, plugins for behavior, skills for instructions. That stack is the default now. Claude Code invented two of the three formats. Omp's stream rules are the oddball: a regex on the token stream aborts mid-sentence, injects a correction, resumes. Project rules without stuffing every prompt. Goose is MCP-purist. Pi ships no MCP in core and hands everything to a TypeScript extension SDK; 67k stars say that resonated.

Git and review workflows

Git used to be an afterthought. Aider fixed that in 2023 with one atomic commit per AI change; every serious tool followed. The 2026 wrinkles look different. Copilot CLI owns the platform — review, PRs, issues are native objects. Omp treats git as another tool surface: omp commit splits unrelated changes into ordered commits, rejects dependency cycles before writing, and turns merge conflicts into files resolved with @ours / @theirs / @base instead of fragile text surgery. Built-in review is baseline now: 4 of the 5 comparison tools ship it; Continue built its CI identity around the same idea.

Safety and trust

There is no clean option. Codex CLI is the only lab agent that's open source, OS-sandboxed, and local-model-capable at once. Claude Code and Copilot sandbox well but stay closed and cloud-bound. OpenCode and Omp are fully open and run local models, but permissions instead of kernel isolation. CodeBuddy sandboxes execution, OpenHands containerizes everything, Amp Free promises no training on your code, OpenCode says it stores no code server-side. Enterprises weight different rows than individual developers — which is half the reason platform CLIs exist.

Automation and surfaces

Labs and platforms win on surfaces — desktop, web, mobile, handoff from phone to laptop. Open tools bet on protocols instead. ACP (Junie, Kimi Code, Mistral Vibe, CodeBuddy, Omp) lets one agent serve any compatible editor. gptme runs persistent background agents. Continue's cn was CI-first before that was trendy. Omp's /collab shares a live session over a link with client-side encryption. Strange feature, useful in practice.

Web, media, and input

An agent that can't fetch docs, check an issue, or poke a browser hands the work back to a human. Codex made browser control a first-class tool in April. Omp chains 18 search providers and runs headless Chromium that can drive a local Electron app. PDFs, SQLite fixtures, notebooks — minor on a checklist, painful when the alternative is copy-pasting spec pages into the prompt. Aider had voice input years ago; Codex takes speech now; Open Interpreter still means "control the whole machine."

Cost engineering

Same refactor prompt, heavy day, Claude Code vs Reasonix: credit meter on one side, 99.8% cache hit rate and a $12 bill on the other. That gap is why cost engineering became a feature.

Most harnesses treat provider-side caching as luck. Reasonix treats it as design — byte-stable context so DeepSeek's prefix cache keeps hitting, 99.82% on heavy days per their reports. Automations too marginal for Claude tokens start to pencil out. Omp routes sub-tasks to cheap models by role and rotates credentials. Subscription agents hide the bill until headless usage gets its own meter, which Claude Code added in June.

Feature leaders

Omp has the widest feature set in the comparison. Tiny bus factor, and co-trained models still win on gnarly tasks, so don't bet the company on it. But when Claude Code ships something new, check Omp first; the community fork often had it earlier.

Money, benchmarks, and trust

The pricing story split in the 1st half of 2026. Lab agents bundle inference into subscriptions — usually the cheapest frontier access, with lock-in attached. BYOK harnesses trade convenience for freedom and can arbitrage open-weight pricing;