Crush Coding Agent Review 2026: CLI, Pricing, Install, and Alternatives

Crush is Charm's terminal-native coding agent. It combines a polished TUI with repository editing, shell tools, model-provider choice, LSP context, MCP integrations, and reusable Agent Skills. It is aimed at developers who want an agent in the terminal without making one model vendor or AI-first IDE the center of their workflow.

The short verdict: Crush is a strong option for terminal users who value provider flexibility and extensibility, but it is not a zero-cost or fully open-source default. The application is free to install, while model access may require a subscription or usage-based API account. Its current FSL-1.1-MIT license is source-available and restricts competing commercial use until each release converts to MIT after two years.

Quick Verdict

Question	Crush answer
What is it?	A terminal coding agent from Charm with a full-screen TUI
Best for	Shell-first developers who want multiple model providers, LSP, MCP, and Agent Skills
Not ideal for	Teams that require an OSI-approved open-source license today or a managed IDE experience
Application price	Free to install and use
Model cost	Charm Hyper or a separate model provider may add subscription or API charges
Current license	FSL-1.1-MIT, a source-available license with a future MIT grant
Latest release checked	v0.85.0, released July 16, 2026
Main risk	A powerful terminal agent can run consequential commands, and provider costs can be unpredictable

What Is the Crush Coding Agent?

Crush is an interactive coding agent that runs in a terminal. You open it in a repository, choose a model, and ask it to inspect files, explain code, make edits, and run commands. The interface maintains project sessions and lets you change models while preserving the conversation context.

It is not an IDE, despite searches such as "crush ide." It does not replace VS Code, JetBrains, or another editor. Crush is better understood as a terminal control surface for agentic repository work, similar in category to Claude Code, OpenCode, Aider, and Goose.

The official project highlights four capabilities that make it more than a basic chat wrapper:

• Multi-model support: use Charm Hyper or connect providers such as Anthropic, OpenAI, Gemini, OpenRouter, Groq, and compatible endpoints.
• LSP context: connect language servers so the agent can use code intelligence alongside file and shell tools.
• MCP integrations: add external tools over stdio, HTTP, or SSE transports.
• Agent Skills: install reusable SKILL.md packages that teach the agent specialized workflows.

Crush can also initialize a repository context file, respect .gitignore and .crushignore, maintain project-local configuration, and ask for permission before tool calls.

Official Links

Use these primary sources instead of similarly named extensions or older package references:

• Official page: charm.land/crush
• GitHub repository: charmbracelet/crush
• Releases: Crush releases
• npm package: @charmland/crush
• License: FSL-1.1-MIT license text

As of July 18, 2026, the public repository showed about 26.6k stars and 2k forks. The latest stable GitHub release and npm package were v0.85.0. Those numbers indicate substantial interest and active maintenance, but they do not prove that Crush will outperform another agent on your repositories.

How to Install Crush

The official README documents several installation paths. On macOS or Linux with Homebrew:

brew install charmbracelet/tap/crush

With npm:

npm install -g @charmland/crush

On Windows with Winget:

winget install charmbracelet.crush

The project also publishes Scoop, Debian, RPM, Nix, BSD, downloadable binary, and Go installation options. After installation, enter a repository and start the TUI:

cd your-project
crush

The simplest setup is to select a model in the model picker and authenticate. Crush promotes Charm Hyper as its official provider, but you can instead connect another supported provider with an API key. Provider and model availability changes quickly, so verify the current README before standardizing a team setup.

Pricing Reality

Crush is free to download and run. That does not make the complete coding workflow free.

You still need model access. The official provider, Charm Hyper, is subscription-based and advertises a free tier. Other options use their own API keys, subscriptions, quotas, and token prices. A long repository task that reads many files, retries commands, and produces large outputs can cost materially more than a short question.

The practical budget depends on:

• which provider and model you select;
• how much repository context enters each turn;
• whether the model supports prompt caching;
• how often the agent retries or revises work;
• whether your provider plan includes the required access.

For cost-sensitive evaluation, start with a small issue, inspect the provider's usage dashboard, and compare the completed result rather than the advertised per-token price alone.

Source-Available, Not Fully Open Source Today

Crush publishes its source code, but the current license is FSL-1.1-MIT. The Functional Source License permits internal use, education, research, modification, and redistribution, but restricts using the software in a competing commercial product or service.

Each version receives an MIT license grant on the second anniversary of its release. That future conversion is useful, but it does not make a new release OSI-approved open source today.

This distinction will not block most individual developers or teams using Crush internally. It can matter for vendors, hosted-service builders, commercial redistributors, and organizations with policies that only approve OSI licenses. Read the upstream license rather than relying on a repository badge or a casual "open source" description.

Workflow Strengths

The terminal experience is unusually polished. Charm maintains widely used terminal libraries and applications, and Crush benefits from that background. Sessions, model selection, permission prompts, and project navigation feel like parts of one product rather than unrelated scripts.

Provider choice reduces lock-in. You can change models without replacing the agent interface. This is valuable when model quality, regional access, latency, and pricing shift faster than team workflows.

LSP and MCP create useful extension paths. LSP can add language-aware context, while MCP connects databases, documentation, issue trackers, or custom tools. Neither guarantees better output, but both give advanced users more control than a fixed tool set.

Configuration can live with the repository. A project can define providers, permissions, ignore rules, LSPs, and MCP servers in crush.json or .crush.json. Crush can also create an AGENTS.md-style context file for durable project guidance.

The release cadence is active. Version 0.85.0 added a first-class question tool and continued reliability and interface work. Frequent releases are a positive maintenance signal, although they also increase the need to check current configuration docs.

Tradeoffs and Risks

Model access remains an external dependency. Crush can support many providers, but outages, rate limits, model removals, and billing changes still affect the workflow.

Terminal permissions require discipline. Crush asks before tool calls by default and supports allowlists. It also offers --yolo, which skips permission prompts. That mode can be convenient in a disposable sandbox and dangerous in a sensitive repository. Use a branch, protect secrets, inspect diffs, and run tests.

Extensibility adds configuration surface. LSP, MCP, hooks, skills, providers, and project instructions can create a capable environment. They can also create a setup that is hard for teammates to reproduce or audit.

The license is a real procurement consideration. Source visibility is valuable, but FSL restrictions are not equivalent to MIT, Apache-2.0, or another OSI license for current releases.

Pseudonymous metrics are enabled by default. The README says prompts and responses are not collected, but usage metadata tied to a device-specific hash is recorded. Users can disable metrics with CRUSH_DISABLE_METRICS=1, DO_NOT_TRACK=1, or the configuration option documented upstream.

Crush vs OpenCode, Claude Code, Aider, and Goose

OpenCode: Compare OpenCode when you want another provider-flexible terminal agent with a strong open ecosystem. Check its current license, provider support, and client architecture because both projects evolve quickly.

Claude Code: Choose Claude Code when Anthropic's models and product integration matter more than provider portability. Crush is the more natural fit when you want one interface across several model vendors.

Aider: Choose Aider when a mature, git-centered workflow and explicit repository mapping are more important than a full-screen TUI. Aider is also a cleaner licensing choice for teams that require a conventional open-source license.

Goose: Consider Goose when you want an extensible open-source agent from Block with a different extension and provider model. Compare the actual approval flow and setup burden on one representative task.

Who Should Use Crush?

Use Crush if:

• you prefer terminal-native development and want a polished interactive interface;
• you need to compare or switch model providers without changing agent tools;
• LSP, MCP, hooks, or Agent Skills are meaningful to your workflow;
• you are comfortable reviewing commands, diffs, tests, and provider usage;
• FSL-1.1-MIT is acceptable for your use case.

Skip or delay it if:

• your organization requires an OSI-approved license for current releases;
• you want a fully managed IDE with enterprise administration and predictable bundled billing;
• you do not want a coding agent to run shell commands in your repository;
• a simpler, git-focused tool already satisfies your needs;
• your team cannot maintain shared provider, permission, and integration configuration.

Bottom Line

Crush is a credible terminal coding agent, not just a visually polished demo. It has active releases, strong public repository traction, broad installation support, provider choice, LSP context, MCP integrations, Agent Skills, and sensible permission controls.

Its strongest advantage is flexibility: the terminal workflow can stay stable while models and integrations change. Its biggest caveats are equally practical: inference still costs money, powerful shell access needs careful review, and FSL-1.1-MIT is source-available rather than fully open source for current releases.

Start with the Crush tool profile, then verify the current official README, latest release, and license before adopting it across a team.