AI Coding Market Map
Agents

CLI Coding Agents

Nascent
Emerging
Growth
Maturing
Mature

Terminal-native AI coding agents that run in your shell. These are the tools developers use to write, refactor, and debug code through conversational commands in the terminal — the most direct path to getting an AI agent working on real code.

Overview Recommendations Trends and Strategic Signals Tools Adoption and Traction New Entrants & Watch List

Overview

Category maturity: Growth. The category is maturing rapidly — two major tools (GitHub Copilot CLI and Kiro CLI) now hold enterprise-ready compliance postures, the Claude Code Analytics API closes a key governance gap, and the OpenCode fork dispute signals that open-source governance is becoming a real concern as adoption scales.

Direction of travel: The category is differentiating along two axes over the next 6-12 months:

  • Depth of enterprise governance — SSO, audit logs, admin policy enforcement, programmatic usage tracking
  • Model specialization — large context reasoning models for complex agentic work vs. fast, cheap coding-specific models for high-frequency tasks

Tools that succeed on both will earn the enterprise standard position; tools that optimize for one will settle into a specialist niche. The Kiro CLI GovCloud launch and the Claude Code Analytics API both point to enterprise governance as the next major competitive front.

Coalesced patterns: The category has converged on a reliable set of foundations:

  • Terminal-native TUI as the primary interface
  • MCP (Model Context Protocol) as the standard tool integration layer
  • Git-native or git-adjacent execution as the baseline workflow
  • Multi-agent coordination as the expected architecture for complex tasks
  • Pricing settled at free individual tiers with per-seat commercial plans for team and enterprise use

Unsolved problems: Governance at scale remains the primary gap:

  • Most tools provide strong individual developer experiences but require significant self-managed effort to deploy safely across an organization: which agents are authorized, how AI-generated code is flagged in version control, and how cost attribution works by team or project
  • Enterprise-grade policy enforcement, centralized agent authorization, and compliance-ready audit trails are still only fully solved by GitHub Copilot CLI; the Claude Code Analytics API is a step forward but not yet complete
  • The OpenCode community fracture surfaces an underappreciated risk: dependency on a fast-moving open-source project carries governance and stability considerations that teams should plan for

Recommendations

1. Start tracking Claude Code adoption metrics before broad rollout. The Analytics API is now available for Team and Enterprise plans, giving admins daily programmatic access to usage metrics by user, model, and cost. Set up tracking now so you have a baseline for measuring productivity impact and managing cost exposure at volume.

2. Evaluate sub-second coding models as a distinct use case. GPT-5.3-Codex-Spark (>1,000 tokens/second) and Gemini 2.5 Flash Lite point to real-time AI coding assistance becoming available within the next 1-2 quarters. Fast models are strongest for high-frequency, lower-complexity tasks like inline suggestions, quick edits, and test generation — a different productivity use case from agentic multi-file work. Plan to evaluate fast models and full reasoning models as distinct tracks rather than treating "which AI coding tool" as a single procurement decision.

1. Amazon Q Developer CLI is now Kiro CLI, and it is closed-source. Starting with version 1.20, AWS rebranded its CLI coding agent from Amazon Q Developer CLI to Kiro CLI. The open-source repository is no longer actively maintained and will receive only critical security fixes. Kiro is now a closed-source product hosted on kiro.dev, powered by Amazon Bedrock, and available in AWS GovCloud (US-East and US-West) regions as of February 2026. This is a significant architecture and procurement shift: teams that evaluated Amazon Q Developer CLI should re-evaluate under the Kiro brand with its updated compliance posture, GovCloud availability, and granular agent steering controls. For PE-backed companies in regulated industries (healthcare, financial services, government), GovCloud availability is a notable differentiator.

2. Amp made a deliberate bet on the terminal and dropped its IDE extensions. Amp's VS Code and Cursor extensions self-destructed on March 5, 2026. The product is now CLI-exclusive, and Amp published a blog post titled "The Coding Agent Is Dead" to frame its view that IDE-integrated agents are a distraction and the terminal is the right surface for serious agentic work. GPT-5.4 is now available in Amp alongside extended thinking. Amp remains a coding agent; the pivot is about surface, not category. Teams that evaluated Amp as an IDE-adjacent tool should revisit their assessment with the CLI-only fit profile in mind.

3. The OpenCode community fractured into two competing projects. A dispute over project ownership between Charmbracelet (Charm) and contributors from SST (Dax and Adam) split OpenCode into two products. Charm maintains Crush (charmbracelet/crush), the original-author fork with a polished TUI, updated March 31, 2026. SST maintains OpenCode (opencode-ai/opencode), the community contributor fork. Both projects are actively developed and share a common lineage but diverging roadmaps. Teams adopting either tool should be explicit about which fork they are standardizing on. Community star counts now aggregate across both repos, making raw GitHub star counts less reliable as a single-project signal.

4. Claude Code Analytics API reached general availability, unlocking programmatic enterprise governance. Anthropic launched the Claude Code Analytics API, giving organization admins programmatic access to daily aggregated usage metrics: sessions, lines of code, commits, pull requests, tool usage, and token and cost data broken down by model and user. Only admins on Team and Enterprise plans can provision API keys. This is a meaningful enterprise-readiness step: organizations can now build custom dashboards, integrate adoption data into engineering metrics platforms, and track cost-per-developer before committing to broad rollout. The API fills a visible gap that enterprise buyers had flagged when comparing Claude Code to GitHub Copilot CLI.

5. Sub-second, real-time coding models are entering the category. OpenAI released GPT-5.3-Codex-Spark as a research preview for ChatGPT Pro users, available in the Codex app, CLI, and IDE extension. Codex-Spark delivers more than 1,000 tokens per second while remaining capable on real-world coding tasks, designed explicitly for a "near-instant" feel. Google added Gemini 2.5 Flash Lite support to Gemini CLI. The direction is clear: model providers are now building distinct SKUs for coding tasks, with smaller, faster, cheaper models targeting high-frequency, lower-complexity coding tasks (autocomplete, inline suggestions, quick edits) alongside the full reasoning models used for multi-file agentic work.

Tools

Claude Code logoClaude Code

  • Maker: Anthropic
  • Model(s): Claude Sonnet 4.6, Claude Opus 4.6 (1M token context window in beta)
  • Access: Subscription via claude.ai; API access via Anthropic Platform; Claude Max, Team, or Enterprise plan required for full agentic use
  • Strengths:
    • Best-in-class multi-file agentic reasoning; strong at complex, long-horizon tasks with 1M token context
    • Analytics API (GA) enables programmatic access to daily usage metrics by user, model, and cost; closes a key enterprise governance gap
    • Agent Teams (multi-agent coordination) shipped in February 2026; Claude Agent SDK exposes the full permissions framework for custom workflows
    • Frequent release cadence: transcript search, PowerShell opt-in preview, subprocess credential scrubbing, and session naming all shipped in Q1 2026
  • Limitations:
    • Centralized admin policy enforcement (organization-wide agent authorization) is less mature than GitHub Copilot CLI and Kiro CLI today; the Analytics API is a strong step, but policy controls are still developing
    • The April 2026 npm packaging incident exposed internal TypeScript source code (no customer data or credentials); the incident points to a maturing DevSecOps practice worth monitoring
    • Opus 4.6 at volume carries meaningful token cost; teams should use the Analytics API to model cost exposure before broad rollout
  • Enterprise readiness: Developing: Analytics API is a meaningful maturity step; team-level cost tracking and admin-level policy enforcement are the next tier
  • Best for: Engineering teams prioritizing agentic quality and complex reasoning; strongest fit for greenfield feature development, large refactors, and teams that want programmatic usage visibility
  • This week: Analytics API reached GA. npm source code exposure incident confirmed (v2.1.88; no sensitive data affected). Transcript search, PowerShell opt-in, and subprocess credential scrubbing shipped.

Codex CLI logoCodex CLI (OpenAI)

  • Maker: OpenAI
  • Model(s): GPT-5.3-Codex (full reasoning); GPT-5.3-Codex-Spark (real-time, research preview); GPT-5.4 available via /model selector
  • Access: Available via npm or Homebrew; included with ChatGPT Pro, Plus, Business, Enterprise, Edu; expanded to Free and Go tiers
  • Strengths:
    • GPT-5.3-Codex-Spark (research preview) delivers more than 1,000 tokens per second, enabling near-instant coding assistance for high-frequency tasks
    • Multi-agent v2 workflows use readable path-based agent addresses (/root/agent_a), making parallel sessions easier to manage and audit
    • Plugins are now a first-class workflow: sync at startup, browse in /plugins, install or remove with guided auth
    • Broad model selection within a single session via /model command; pipe/stdin workflow support for scripting integration
  • Limitations:
    • Enterprise-grade audit logging and admin policy enforcement are developing; GitHub Copilot CLI and Kiro CLI are ahead on organizational governance today
    • GPT-5.3-Codex-Spark is research preview; teams should treat it as an evaluation track rather than production standard until GA
    • Rate limits vary meaningfully by plan tier; confirm limits against anticipated usage volume before committing
  • Enterprise readiness: Developing: OpenAI Business and Enterprise agreements provide data privacy protections; admin controls and organization-level policy enforcement are on the roadmap
  • Best for: Teams in the OpenAI ecosystem; strong for high-frequency coding tasks (with Spark) and research-plus-code hybrid workflows (with full Codex)
  • This week: GPT-5.3-Codex-Spark launched in research preview (>1,000 tokens/second). Multi-agent v2 path-based addressing shipped. Plugin workflow promoted to first-class feature.

GitHub Copilot CLI logoGitHub Copilot CLI

  • Maker: Microsoft / GitHub
  • Model(s): GPT-4o, Claude Sonnet (Anthropic), Gemini (multi-model support)
  • Access: Included with all Copilot subscription tiers (Individual, Business, Enterprise); Enterprise plan required for organization-level admin controls
  • Strengths:
    • The only tool in the category with fully mature organization-level admin controls: administrators define authorized agents, models, and policies across the org
    • MCP server built in; custom MCP servers supported; v1.0.12 fixed workspace MCP server loading and plugin hook environment variables
    • GA since February 2026; procurement-ready under Microsoft enterprise agreements with SLA, compliance documentation, and data residency options
    • Deep GitHub native context (PRs, issues, repos) available in every agent session without manual setup
  • Limitations:
    • Full value requires GitHub Enterprise; teams on GitLab, Bitbucket, or custom VCS should verify integration depth before committing
    • Agentic task quality on complex multi-file work is still maturing relative to Claude Code and Codex CLI on pure coding benchmarks
    • Multi-model breadth means teams will want to evaluate which model to pin for their specific workloads; "best for everything" claims should be tested
  • Enterprise readiness: Enterprise-ready: SOC 2, audit logs, organization-level admin controls, SSO via GitHub Enterprise, Microsoft data residency options, SLA support
  • Best for: Teams on GitHub Enterprise seeking the lowest-friction, compliance-ready path to team-wide agentic coding adoption
  • This week: v1.0.14 released March 31, 2026 (Anthropic BYOM image handling corrected, MCP server startup improved, Shift+Enter newline support in Kitty terminals). No breaking changes.

Kiro CLI logoKiro CLI (formerly Amazon Q Developer CLI)

  • Maker: Amazon Web Services
  • Model(s): Amazon Bedrock foundation models; supports multiple model providers via Bedrock
  • Access: Closed-source; free tier available; Pro tier at $19/user/month; integrated into AWS accounts; available on kiro.dev
  • Strengths:
    • Full AWS GovCloud (US-East and US-West) availability as of February 2026; meaningful for regulated industries and public sector customers
    • Native AWS integration: reads and writes files, calls AWS APIs, runs bash commands; spec-driven development breaks prompts into logical implementation steps with tests and API integrations generated together
    • Covered under AWS enterprise agreements with IAM policy integration and AWS compliance certifications (SOC 2, ISO 27001, HIPAA eligible)
    • Granular web fetch URL controls and keyboard shortcuts for custom agents added in recent releases; MCP support native
  • Limitations:
    • The open-source Q Developer CLI is no longer actively maintained; teams that had evaluated the open-source version should re-assess against the closed-source Kiro product
    • Strongest on AWS-adjacent tasks; teams evaluating it for general-purpose software development will find Claude Code or Codex CLI more capable on that dimension
    • Differentiation is clearest when AWS is central to the stack; teams on other clouds will find less leverage
  • Enterprise readiness: Enterprise-ready: AWS GovCloud availability, IAM integration, AWS compliance certifications, enterprise support tiers, audit trail via AWS CloudTrail
  • Best for: Engineering teams with AWS-centric infrastructure, particularly those in regulated industries requiring GovCloud compliance
  • This week: AWS GovCloud regional availability confirmed. Rebrand from Amazon Q Developer CLI to Kiro CLI complete (closed-source as of v1.20). Open-source repository entering maintenance-only mode.

Gemini CLI logoGemini CLI (Google)

  • Maker: Google
  • Model(s): Gemini 2.5 Pro (primary); Gemini 2.5 Flash Lite (newly added); 1M token context window
  • Access: Open-source (Apache 2.0); free tier via Gemini API; paid via Google Cloud / Gemini Code Assist; higher quotas for Google AI Pro and AI Ultra subscribers
  • Strengths:
    • Gemini 2.5 Flash Lite support added; enables teams to route lower-complexity tasks to a faster, cheaper model within the same tool
    • Experimental memory manager agent replaces the save_memory tool; a step toward persistent project context across sessions
    • Built-in sandboxing (Linux bubblewrap/seccomp) for tool isolation; plan mode provides a read-only analysis phase before any file writes
    • Fully open source with MCP support and ACP SDK 0.16.1 for extended integrations
  • Limitations:
    • Centralized admin tooling for enterprise engineering orgs is still maturing; GitHub Copilot CLI and Kiro CLI are ahead on organizational governance controls
    • Memory manager is experimental; teams that want stable, persistent project context should plan for iteration
    • The community and third-party integration ecosystem is growing but still smaller than Aider and OpenCode (SST) at this stage
  • Enterprise readiness: Developing: strong technical security controls (sandboxing, plan mode) and Google Cloud compliance posture on paid tiers; centralized admin tooling is the next maturity step
  • Best for: Google Cloud-centric teams; teams that want a large context window, a read-before-write safety model, and built-in sandbox isolation
  • This week: Gemini 2.5 Flash Lite support added. Experimental memory manager agent available. ACP SDK upgraded to 0.16.1. Higher quota limits for premium Google AI subscribers.

Aider logoAider

  • Maker: Open-source (Paul Gauthier, community-maintained)
  • Model(s): Model-agnostic; supports Claude, GPT-4o, Gemini, and local models via Ollama
  • Access: Fully open-source (Apache 2.0); free; community-supported
  • Strengths:
    • Most established open-source CLI coding agent with the deepest install base: 4.1M installs, 15B tokens/week in production
    • Git-native by design: every change is committed with an explanatory message, providing strong auditability out of the box
    • Model-agnostic; works with local models for air-gapped environments where code cannot leave the network
    • Zero license cost with no per-seat pricing; the lowest total cost of ownership in the category
  • Limitations:
    • Suited to individual or small-team use; enterprise deployment runs through self-managed governance since vendor-provided SSO, SLA, and admin controls are not part of the offering
    • Multi-agent and parallel task support is minimal compared to commercial alternatives; single-agent architecture is the current design
    • UX rewards teams willing to invest in configuration; the tradeoff for maximum model flexibility
  • Enterprise readiness: Early: well-suited for individual developer adoption; organization-wide governance runs through self-managed policies rather than vendor controls
  • Best for: Individual developers or small teams wanting maximum model flexibility at zero license cost; teams in regulated industries who need local model execution
  • This week: No new signals. 41.6K GitHub stars current.

OpenCode logoOpenCode (SST fork)

  • Maker: Open-source (SST; contributors Dax and Adam, community-maintained)
  • Model(s): Supports 75+ LLM providers including Claude, GPT, Gemini, and local models
  • Access: Open-source; free; community-supported; $10/month paid tier available
  • Strengths:
    • 134K GitHub stars as of April 2026, the highest star count of any open-source tool in the category
    • 2.5M monthly developers; the broadest provider support in the open-source segment (75+ LLMs)
    • v1.3.10 released recently with storage migration reliability improvements and Catppuccin theme contrast fixes
    • GitHub officially supports using Copilot Pro, Pro+, Business, and Enterprise subscriptions with OpenCode, giving enterprise teams a supported path to standardized credentials
  • Limitations:
    • The Charm/SST fork dispute is recent; the project is actively maintained under SST but teams should understand they are depending on a community fork, not the original authored project
    • Suited to individual or team use; enterprise deployment works best with self-managed governance; vendor-provided SSO, audit logs, and compliance documentation are not yet part of the offering
    • Rapid growth means the codebase and community are still stabilizing; track releases closely in production deployments
  • Enterprise readiness: Early: GitHub Copilot credential integration is a promising enterprise bridge; dedicated enterprise governance features are on a future roadmap
  • Best for: Teams wanting maximum LLM provider flexibility at minimal cost; developers who want to switch underlying models freely without vendor lock-in
  • This week: 134K GitHub stars (up from 95K at prior run). v1.3.10 released (storage migration reliability, theme improvements). Community governance context: SST fork now operates independently of Charmbracelet's Crush.

Amp logoAmp

  • Maker: Amp (independent company, spun out of Sourcegraph December 2025; backed by Craft, Redpoint, Sequoia, a16z, Goldcrest)
  • Model(s): GPT-5.4 (current primary); extended thinking support; model selection built in
  • Access: CLI-only as of March 5, 2026; research preview; pricing not publicly disclosed
  • Strengths:
    • CLI is now the primary and exclusive interface: focused, portable, runs anywhere, and attracts advanced terminal-native developers
    • GPT-5.4 now available with extended thinking for deeper reasoning on complex tasks
    • Unconstrained token usage design: uses the best model for each task rather than capping for cost
    • Free tier ad-removal signals a cleaner, developer-first experience
  • Limitations:
    • VS Code and Cursor extensions are permanently deprecated as of March 5; teams that adopted Amp for IDE integration need to migrate to CLI-only workflows
    • Still in research preview; enterprise features, pricing, and SLA remain undisclosed; procurement requires waiting for GA
    • The "coding agent is dead" messaging is a positioning claim about IDE-first agents, not a category exit; the product remains a coding agent, and teams should track whether the CLI-only focus produces a clearer enterprise roadmap
  • Enterprise readiness: Early: research preview with strong investor backing; CLI-only pivot simplifies the deployment surface but enterprise governance is still on a future roadmap
  • Best for: Advanced terminal-native developers and engineering teams willing to run cutting-edge tooling; best as a quality benchmark or individual power-user tool rather than a broad org rollout
  • This week: VS Code and Cursor extensions permanently deprecated (March 5). CLI is now the sole interface. GPT-5.4 available. Free tier removes ads. Blog post "The Coding Agent Is Dead" published.

mini-SWE-agent logomini-SWE-agent (Princeton / Stanford)

  • Maker: Princeton NLP Group and Stanford researchers (SWE-agent organization)
  • Model(s): Model-agnostic via LiteLLM, OpenRouter, and Portkey; works with Claude, GPT-4o, Gemini, and open-source models
  • Access: Free, open-source (MIT license); install via pip. github.com/SWE-agent/mini-swe-agent
  • Strengths:
    • Scores greater than 74% on SWE-bench Verified in approximately 100 lines of Python — the highest benchmark-to-code-size ratio of any agent in the category
    • Adopted by Meta, NVIDIA, IBM, Essential AI, Nebius, and Anyscale, providing production-use validation beyond academic benchmarks
    • Radically simple architecture: relies solely on bash rather than specialized tools; no large configurations, no monorepo. The entire core agent class is ~100 lines
    • The recommended successor to the original SWE-agent; the maintainers explicitly direct users to mini-swe-agent going forward
    • Multiple sandboxing backends (Docker, Podman, Singularity/Apptainer, Bubblewrap) give teams flexibility over isolation without locking into a single container runtime
  • Limitations:
    • Designed as a foundation for teams willing to customize and extend; not a turnkey product with a polished UX or managed service
    • Research-team support model; teams seeking SLA-backed vendor support should evaluate commercial alternatives first
    • Benchmark scores are measured on SWE-bench, a curated set of real GitHub issues; real-world performance on a specific codebase will vary and should be validated with a pilot
  • Enterprise readiness: Early. Open-source with confirmed enterprise adopters; no managed SaaS or enterprise SLA. Data residency is fully controlled by the team's own execution environment.
  • Best for: Engineering teams that want a high-performance, model-agnostic coding agent as a foundation for custom internal tooling, particularly those comfortable building on open-source research infrastructure.
  • This week: No new release signals. The project continues active development as the successor to SWE-agent.

Adoption and Traction

  • OpenCode (SST fork): 134K GitHub stars as of early April 2026, up from 95K at prior run. Community growth rate remains the highest in the open-source segment, despite the Charm/Crush fork.
  • Crush (Charmbracelet): New independent project as of July 2025 rebrand; active development, updated March 31, 2026. Multi-model, LSP-integrated, project-isolated. No star count signal yet from separate repo.
  • Claude Code: Analytics API launched (programmatic daily usage metrics via API). npm source code exposure incident (v2.1.88, 57MB map file of TypeScript source) confirmed by Anthropic; no customer data or credentials affected.
  • Codex CLI (OpenAI): GPT-5.3-Codex-Spark research preview launched (>1,000 tokens/second, real-time coding focus). Multi-agent v2 workflows now use readable path-based agent addresses. Codex app on Windows now stable.
  • GitHub Copilot CLI: v1.0.14 released March 31, 2026 (Anthropic BYOM image handling fix, MCP server startup improvements, Shift+Enter newline in Kitty terminals). No breaking changes.
  • Gemini CLI: Gemini 2.5 Flash Lite support added. Experimental memory manager agent available (replacing save_memory tool). ACP SDK upgraded from 0.12 to 0.16.1. Higher quota limits for Google AI Pro and AI Ultra subscribers.
  • Kiro CLI (Amazon): AWS GovCloud (US) regional availability confirmed. Granular web fetch URL controls, keyboard shortcuts for custom agents, enhanced diff views. Closed-source as of v1.20.
  • Amp: CLI-only as of March 5 (VS Code and Cursor extensions deprecated). GPT-5.4 now available in Amp. Free tier removes ads.

New Entrants & Watch List

Open SWE (LangChain) Launched March 17, 2026. An open-source framework that captures the patterns used by Stripe, Ramp, and Coinbase for their internal coding agents. Built on Deep Agents and LangGraph. A CLI version is in development for local execution with more agentic autonomy. Worth watching because it provides architectural patterns rather than a finished product: engineering teams at growth-stage companies that want to build internal coding agents rather than buy off-the-shelf tools now have a reference architecture from practitioners at scale.

Next → IDE Agents Agents