What it does
6 features
Crush · Terminal coding agent
Product and access
Crush is developed by Charmbracelet, Inc. Charm also runs its own model provider, Hyper, for Crush, which the documentation calls the official provider.
Core features
6 features from provider sources · checked on 5 November 2026. Availability and limits are directly below.
Multiple models and providers
You choose models from various providers or add your own through OpenAI- or Anthropic-compatible APIs. The model can be switched mid-session, and the README says context is preserved.
Source 1LSP context
Crush uses language servers for additional context; you set them up with a command.
Source 1MCP extensions
MCP servers over stdio, http and sse extend the tools. HTTP and SSE servers have built-in OAuth support.
Source 1Permission prompts for tool calls
By default Crush asks before tool calls. You can allow or deny tools, or skip all prompts with --yolo.
Source 1Hooks and Agent Skills
Hooks are shell commands at points in the agent run; they can block, allow or modify tool calls. Skills follow the open Agent Skills standard.
Source 3
Usable AI models
Open choice
Crush works with your own keys for many providers such as Anthropic, OpenAI, Gemini, Z.ai or Moonshot, with OpenRouter and Charm's Hyper service. You can add further providers yourself through OpenAI- or Anthropic-compatible interfaces. Crush discovers local models from Ollama, LM Studio, llama.cpp and oMLX.
Model families named
- Anthropic (Claude)
- OpenAI (GPT)
- Google (Gemini)
- Z.ai (GLM)
- MiniMax
- Moonshot (Kimi)
- xAI (Grok)
- DeepSeek
- Other models can be connected freely
- Model choice
- You choose
- Your own API key
- Yes
- Router or custom endpoint
- Yes
- Local models
- Yes
The rating covers the generally available main access. Documented · Sources accessed on
Does it fit my task?
2 practical tasks · 5 requirements and limits
A good fit if …
- you work in the terminal and want to use models from various providers or local models
- you want to approve tool calls before they run and need hooks for your own rules
- you want to connect LSP context and MCP servers in a terminal agent
Less suitable if …
- you need an ordinary open-source licence or are building a competing commercial product on it
- you expect schedules or event-driven unattended runs
- you need central team and admin management
Editorial assessment based on provider documentation. Not a product test.
Practical tasks
Editorial examples based on documented features. No product tests performed.
Code change with approval
- Your input
- A project folder with AGENTS.md and a limited change.
- Possible result
- The agent proposes file changes and commands and runs them after your approval.
- What to check
- Check the diff and tests; grant lasting permissions only selectively.
Enforce rules through hooks
- Your input
- A hook per event, such as before tool calls, with your own script.
- Possible result
- The script can block or allow calls or modify inputs.
- What to check
- Test the hook scripts; on conflict, the docs say a deny beats an allow.
Start Crush in a project folder and have it make a limited code change.
Your possible prompt
Read the project rules. Investigate [error], propose a small change, and run the relevant checks. Show me the diff and any remaining uncertainties.
Adjust the task, data and permissions.
Requirements and limits
Requirement
Model access required
You need a Hyper account, a provider API key or a local model.
Source 1Limitation
Licence with competing-use restriction
FSL-1.1-MIT permits any purpose other than a competing use, explicitly including internal use, non-commercial education and research, and professional services for a licensee. A competing use is a commercial product or service that substitutes for Crush or another Charm offering using it, or provides substantially similar functionality. Two years after a version is released, the MIT licence applies to it.
Source 2Limitation
Configuration is executable code
crushrc and crush.json are treated as trusted code and can run commands. The README advises against launching Crush in folders whose configuration you have not reviewed.
Source 1Limitation
--yolo skips all prompts
The flag turns off permission prompts; the README explicitly warns about it.
Source 1Not yet confirmed
Schedules and team administration
The reviewed pages describe neither schedules nor central administration for organisations.
Source 1
Can I use it?
Availability still unclear
Europe & regional availability
README and licence reviewed. They name no country release or data region; only Hyper is described as GDPR compliant.
Reviewed scope: Local software; model providers separately
Reviewed on:
Availability still unclear
Europe: no availability statement
Crush runs locally on your machine. The README calls Hyper GDPR compliant and says it offers zero data retention; it names no countries or data regions. Pseudonymous usage metrics are on and can be turned off with CRUSH_DISABLE_METRICS=1 or DO_NOT_TRACK=1. Check model providers separately.
Region, reason and source
Region covered: EU/EEA, United Kingdom and Switzerland
Documented reason: The reviewed sources name no country release for this.
Accounts & permissions
Based on provider documentation. Permissions depend on the account, integration and organization policies. Review the permissions actually requested before use.
Configuration covered: Local coding agent with your own model access
Sources reviewed:
Show all 3 points
Open in reviewed sources
Admin approval
The reviewed pages do not mention approval or restriction by organisation administrators.
Evidence
Memory and conversation continuity
1 of 6 criteria documented
Sessions are organised per project; you control context files such as AGENTS.md. The reviewed pages do not say where histories are stored or how to delete them.
Sessions and context files of the coding agent, per project
Sources accessed:
Documentation by criterion
StoragePartly documented
Storage and scope
The README mentions multiple sessions per project and a local database. Application state is under ~/.local/share/crush, as JSON only. It does not say where the conversation history is kept.
SelectionPartly documented
Selection and timing
Initialisation creates a context file, by default AGENTS.md, with build commands and conventions. Global rules live in CRUSH.md or AGENTS.md. The configuration docs list an option to summarise long conversations automatically.
Edit & deletePartly documented
Inspect, edit and delete
You edit context files yourself, .crushignore excludes files, and logs are in ./.crush/logs/crush.log. The reviewed pages do not describe deleting sessions.
Access boundariesDocumented
Agent and channel boundaries
Sessions belong to a project. Multiple clients with the same working directory share session list, history, permission queue, LSP and MCP state. Global context files apply across projects.
RetrievalPartly documented
Retrieval
The session picker gives access to existing sessions; context is preserved when switching models. The pages describe no search across earlier sessions for automatic retrieval.
New conversationsPartly documented
Session reset and compaction
Each new invocation starts in a fresh session by default. The hooks docs list events before and after compaction; PreCompact can block it.
Cost and effort
3 plans · Prices as of 5 November 2026
Prices and included usage
Pricing checked
The reviewed pages state no price for the software. You pay for the model depending on route: your own key, or Hyper with a free plan, paid plans and prepaid bundles.
| Plan | Price | Included |
|---|---|---|
| Crush | No software fee stated | Coding agent under FSL-1.1-MIT, with your own model access. Price source |
| Hyper Free | Free | 100 Hypercredits per month for every account. Price source |
| Hyper Paid / Prepaid | Prices not stated | According to the docs, higher monthly allowances and priority access to models; prepaid bundles do not expire. Price source |
- The hyper.charm.land pages could not be read; Hyper prices and privacy details remain unchecked.
Setup
Install via Homebrew, npm, a package manager or go install, then choose a model: a Hyper account or your own API key.
Operation & oversight
Based on public provider documentation. Assessments are editorial inferences from the stated properties. Actual time requirements and reliability have not been measured.
Deployment covered: Local operation in the terminal
Sources reviewed:
Editorial assessment
You control permissions per tool and keep the prompt by default. Configuration files are executable code, and the reviewed pages are silent on schedules and crashes.
Show all 6 points
Documented with conditions
Files and working state
Crush keeps sessions per project. State is under ~/.local/share/crush; the README does not say exactly where the history is.
EvidenceDocumented
Sign-ins
Sign in for Hyper in the model picker, otherwise API keys or environment variables. For Bedrock an AWS login also applies. The pages name no durations, but AWS SSO sessions expire periodically and can be refreshed with aws_auth_refresh.
EvidenceOpen in reviewed sources
Starting further runs
You start Crush in the terminal and follow the run. The README mentions a shared backend (crush serve) for several interfaces, but no schedules or triggers.
EvidenceDocumented with conditions
Detecting errors
Logs are in ./.crush/logs/crush.log, readable with crush logs and more verbose with --debug. The hooks docs list a StopFailure event. Apart from the Bedrock credential refresh through aws_auth_refresh, the pages describe no automatic retries.
Evidence · EvidenceDocumented with conditions
Resuming after failures
On a shared backend, each new invocation starts a fresh session by default, and the session picker joins one another client has open. The README describes no crash recovery.
Evidence
Evidence and open questions
6 sources · 2 open points
Open points
The reviewed sources contain no reliable information on these points. A gap says nothing about the product.
Europe: Availability still unclear
Requirements and limits: not yet confirmed
Sources and evidence
Repository, README, licence text, hooks and configuration docs and two Hyper docs pages reviewed. Feature descriptions come from the provider; example tasks are editorial. Hyper plan prices and the hyper.charm.land pages could not be read. Hands-on tests are still pending.
Editorial team: Michael Meierhoff · sources reviewed on . Provider information is not independent evidence of performance.
Provider
Charmbracelet, Inc. publishes Crush in its own GitHub repository; the licence names the company as licensor.
- Operator
- Charmbracelet, Inc.
- Organization
- Provider
- Website
- github.com
- Repository
- charmbracelet/crush
- GitHub stars
- 28.966 · Reviewed
- License
- FSL-1.1-MIT
Changes to this profile 1
A terminal coding agent from Charm with several model providers.
These notes describe editorial changes. Source and pricing review dates appear in their respective sections.