README.md · crush · git.iko.soy

git.iko.soy/crush

crush/README.md

914 lines 28 kB history raw

1# Crush

3

4 <a href="https://stuff.charm.sh/crush/charm-crush.png"><img width="450" alt="Charm Crush Logo" src="https://github.com/user-attachments/assets/cf8ca3ce-8b02-43f0-9d0f-5a331488da4b" /></a>

5 <a href="https://github.com/charmbracelet/crush/releases"><img src="https://img.shields.io/github/release/charmbracelet/crush" alt="Latest Release"></a>

6 <a href="https://github.com/charmbracelet/crush/actions"><img src="https://github.com/charmbracelet/crush/actions/workflows/build.yml/badge.svg" alt="Build Status"></a>

7

9Your new coding bestie, now available in your favourite terminal. Your tools, your code, and your workflows, wired into your LLM of choice.

10终端里的编程新搭档， 无缝接入你的工具、代码与工作流，全面兼容主流 LLM 模型。

12<img width="800" alt="Crush Demo" src="https://github.com/user-attachments/assets/58280caf-851b-470a-b6f7-d5c4ea8a1968" />

14## Features

16- **Multi-Model:** choose from a wide range of LLMs or add your own via OpenAI- or Anthropic-compatible APIs

17- **Flexible:** switch LLMs mid-session while preserving context

18- **Session-Based:** maintain multiple work sessions and contexts per project

19- **LSP-Enhanced:** Crush uses LSPs for additional context, just like you do

20- **Extensible:** add capabilities via MCPs (`http`, `stdio`, and `sse`)

21- **Works Everywhere:** first-class support in every terminal on macOS, Linux, Windows (PowerShell and WSL), Android, FreeBSD, OpenBSD, and NetBSD

22- **Industrial Grade:** built on the Charm ecosystem, powering 25k+ applications, from leading open source projects to business-critical infrastructure

24## Installation

26Use a package manager:

28```bash

29# Homebrew

30brew install charmbracelet/tap/crush

32# NPM

33npm install -g @charmland/crush

35# Arch Linux (btw)

36yay -S crush-bin

38# Nix

39nix run github:numtide/nix-ai-tools#crush

41# FreeBSD

42pkg install crush

43```

45Windows users:

47```bash

48# Winget

49winget install charmbracelet.crush

51# Scoop

52scoop bucket add charm https://github.com/charmbracelet/scoop-bucket.git

53scoop install crush

54```

56<details>

57<summary>Nix (NUR)</summary>

59Crush is available via the official Charm [NUR](https://github.com/nix-community/NUR) in `nur.repos.charmbracelet.crush`, which is the most up-to-date way to get Crush in Nix.

61You can also try out Crush via the NUR with `nix-shell`:

63```bash

64# Add the NUR channel.

65nix-channel --add https://github.com/nix-community/NUR/archive/main.tar.gz nur

66nix-channel --update

68# Get Crush in a Nix shell.

69nix-shell -p '(import <nur> { pkgs = import <nixpkgs> {}; }).repos.charmbracelet.crush'

70```

72### NixOS & Home Manager Module Usage via NUR

74Crush provides NixOS and Home Manager modules via NUR.

75You can use these modules directly in your flake by importing them from NUR. Since it auto detects whether its a home manager or nixos context you can use the import the exact same way :)

77```nix

78{

79 inputs = {

80 nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";

81 nur.url = "github:nix-community/NUR";

82 };

84 outputs = { self, nixpkgs, nur, ... }: {

85 nixosConfigurations.your-hostname = nixpkgs.lib.nixosSystem {

86 system = "x86_64-linux";

87 modules = [

88 nur.modules.nixos.default

89 nur.repos.charmbracelet.modules.crush

90 {

91 programs.crush = {

92 enable = true;

93 settings = {

94 providers = {

95 openai = {

96 id = "openai";

97 name = "OpenAI";

98 base_url = "https://api.openai.com/v1";

99 type = "openai";

100 api_key = "sk-fake123456789abcdef...";

101 models = [

102 {

103 id = "gpt-4";

104 name = "GPT-4";

105 }

106 ];

107 };

108 };

109 lsp = {

110 go = { command = "gopls"; enabled = true; };

111 nix = { command = "nil"; enabled = true; };

112 };

113 options = {

114 context_paths = [ "/etc/nixos/configuration.nix" ];

115 tui = { compact_mode = true; };

116 debug = false;

117 };

118 };

119 };

120 }

121 ];

122 };

123 };

124}

125```

126

127</details>

128

129<details>

130<summary>Debian/Ubuntu</summary>

131

132```bash

133sudo mkdir -p /etc/apt/keyrings

134curl -fsSL https://repo.charm.sh/apt/gpg.key | sudo gpg --dearmor -o /etc/apt/keyrings/charm.gpg

135echo "deb [signed-by=/etc/apt/keyrings/charm.gpg] https://repo.charm.sh/apt/ * *" | sudo tee /etc/apt/sources.list.d/charm.list

136sudo apt update && sudo apt install crush

137```

138

139</details>

140

141<details>

142<summary>Fedora/RHEL</summary>

143

144```bash

145echo '[charm]

146name=Charm

147baseurl=https://repo.charm.sh/yum/

148enabled=1

149gpgcheck=1

150gpgkey=https://repo.charm.sh/yum/gpg.key' | sudo tee /etc/yum.repos.d/charm.repo

151sudo yum install crush

152```

153

154</details>

155

156Or, download it:

157

158- [Packages][releases] are available in Debian and RPM formats

159- [Binaries][releases] are available for Linux, macOS, Windows, FreeBSD, OpenBSD, and NetBSD

160

161[releases]: https://github.com/charmbracelet/crush/releases

162

163Or just install it with Go:

164

165```

166go install github.com/charmbracelet/crush@latest

167```

168

169> [!WARNING]

170> Productivity may increase when using Crush and you may find yourself nerd

171> sniped when first using the application. If the symptoms persist, join the

172> [Slack][slack] or [Discord][discord] and nerd snipe the rest of us.

173

174## Getting Started

175

176The quickest way to get started is to grab an API key for your preferred

177provider such as Anthropic, OpenAI, Groq, OpenRouter, or Vercel AI Gateway and just start

178Crush. You'll be prompted to enter your API key.

179

180That said, you can also set environment variables for preferred providers.

181

182| Environment Variable | Provider |

183| --------------------------- | -------------------------------------------------- |

184| `HYPER_API_KEY` | Charm Hyper |

185| `ANTHROPIC_API_KEY` | Anthropic |

186| `OPENAI_API_KEY` | OpenAI |

187| `VERCEL_API_KEY` | Vercel AI Gateway |

188| `GEMINI_API_KEY` | Google Gemini |

189| `SYNTHETIC_API_KEY` | Synthetic |

190| `ZAI_API_KEY` | Z.ai |

191| `MINIMAX_API_KEY` | MiniMax |

192| `HF_TOKEN` | Hugging Face Inference |

193| `CEREBRAS_API_KEY` | Cerebras |

194| `OPENROUTER_API_KEY` | OpenRouter |

195| `IONET_API_KEY` | io.net |

196| `ALIBABA_SINGAPORE_API_KEY` | Alibaba (Singapore) |

197| `GROQ_API_KEY` | Groq |

198| `AVIAN_API_KEY` | Avian |

199| `OPENCODE_API_KEY` | OpenCode Zen & Go |

200| `VERTEXAI_PROJECT` | Google Cloud VertexAI (Gemini) |

201| `VERTEXAI_LOCATION` | Google Cloud VertexAI (Gemini) |

202| `AWS_ACCESS_KEY_ID` | Amazon Bedrock (Claude) |

203| `AWS_SECRET_ACCESS_KEY` | Amazon Bedrock (Claude) |

204| `AWS_REGION` | Amazon Bedrock (Claude) |

205| `AWS_PROFILE` | Amazon Bedrock (Custom Profile) |

206| `AWS_BEARER_TOKEN_BEDROCK` | Amazon Bedrock |

207| `AZURE_OPENAI_API_ENDPOINT` | Azure OpenAI models |

208| `AZURE_OPENAI_API_KEY` | Azure OpenAI models (optional when using Entra ID) |

209| `AZURE_OPENAI_API_VERSION` | Azure OpenAI models |

210

211### Subscriptions

212

213If you prefer subscription-based usage, here are some plans that work well in

214Crush:

215

216- [Synthetic](https://synthetic.new/pricing)

217- [GLM Coding Plan](https://z.ai/subscribe)

218- [Kimi Code](https://www.kimi.com/membership/pricing)

219- [MiniMax Coding Plan](https://platform.minimax.io/subscribe/coding-plan)

220

221### By the Way

222

223Is there a provider you’d like to see in Crush? Is there an existing model that needs an update?

224

225Crush’s default model listing is managed in [Catwalk](https://github.com/charmbracelet/catwalk), a community-supported, open source repository of Crush-compatible models, and you’re welcome to contribute.

226

227<a href="https://github.com/charmbracelet/catwalk"><img width="174" height="174" alt="Catwalk Badge" src="https://github.com/user-attachments/assets/95b49515-fe82-4409-b10d-5beb0873787d" /></a>

228

229## Configuration

230

231> [!TIP]

232> Crush ships with a builtin `crush-config` skill for configuring itself. In

233> many cases you can simply ask Crush to configure itself.

234

235Crush runs great with no configuration. That said, if you do need or want to

236customize Crush, configuration can be added either local to the project itself,

237or globally, with the following priority:

238

2391. `.crush.json`

2402. `crush.json`

2413. `$HOME/.config/crush/crush.json`

242

243Configuration itself is stored as a JSON object:

244

245```json

246{

247 "this-setting": { "this": "that" },

248 "that-setting": ["ceci", "cela"]

249}

250```

251

252As an additional note, Crush also stores ephemeral data, such as application

253state, in one additional location:

254

255```bash

256# Unix

257$HOME/.local/share/crush/crush.json

258

259# Windows

260%LOCALAPPDATA%\crush\crush.json

261```

262

263> [!TIP]

264> You can override the user and data config locations by setting:

265>

266> - `CRUSH_GLOBAL_CONFIG`

267> - `CRUSH_GLOBAL_DATA`

268

269### LSPs

270

271Crush can use LSPs for additional context to help inform its decisions, just

272like you would. LSPs can be added manually like so:

273

274```json

275{

276 "$schema": "https://charm.land/crush.json",

277 "lsp": {

278 "go": {

279 "command": "gopls",

280 "env": {

281 "GOTOOLCHAIN": "go1.24.5"

282 }

283 },

284 "typescript": {

285 "command": "typescript-language-server",

286 "args": ["--stdio"]

287 },

288 "nix": {

289 "command": "nil"

290 }

291 }

292}

293```

294

295### MCPs

296

297Crush also supports Model Context Protocol (MCP) servers through three transport

298types: `stdio` for command-line servers, `http` for HTTP endpoints, and `sse`

299for Server-Sent Events.

300

301Shell-style value expansion (`$VAR`, `${VAR:-default}`, `$(command)`, quoting,

302nesting) works in `command`, `args`, `env`, `headers`, and `url`, so

303file-based secrets work out of the box. You can use values like `"$TOKEN"`

304or `"$(cat /path/to/secret/token)"`. Expansion runs through Crush's embedded

305shell, so the same syntax works on every supported system, Windows included.

306

307Unset variables expand to the empty string by default, matching bash. For

308required credentials, use `${VAR:?message}` so an unset variable fails loudly

309at load time with `message` instead of silently resolving to empty:

310

311```json

312{ "api_key": "${CODEBERG_TOKEN:?set CODEBERG_TOKEN}" }

313```

314

315Headers (both MCP `headers` and provider `extra_headers`) whose value

316resolves to the empty string are dropped from the outgoing request rather

317than sent as `Header:`. That keeps optional env-gated headers like

318`"OpenAI-Organization": "$OPENAI_ORG_ID"` clean when the variable is unset.

319

320Provider `extra_body` is a non-expanding JSON passthrough; put env-driven

321values in `extra_headers` or the provider's `api_key` / `base_url`, all of

322which do expand.

323

324> **Security note:** `crush.json` is trusted code. Any `$(...)` in it runs at

325> load time with your shell's privileges, before the UI appears. Don't launch

326> Crush in a directory whose `crush.json` you haven't reviewed.

327

328```json

329{

330 "$schema": "https://charm.land/crush.json",

331 "mcp": {

332 "filesystem": {

333 "type": "stdio",

334 "command": "node",

335 "args": ["/path/to/mcp-server.js"],

336 "timeout": 120,

337 "disabled": false,

338 "disabled_tools": ["some-tool-name"],

339 "env": {

340 "NODE_ENV": "production"

341 }

342 },

343 "github": {

344 "type": "http",

345 "url": "https://api.githubcopilot.com/mcp/",

346 "timeout": 120,

347 "disabled": false,

348 "disabled_tools": ["create_issue", "create_pull_request"],

349 "headers": {

350 "Authorization": "Bearer $GH_PAT"

351 }

352 },

353 "streaming-service": {

354 "type": "sse",

355 "url": "https://example.com/mcp/sse",

356 "timeout": 120,

357 "disabled": false,

358 "headers": {

359 "API-Key": "$(echo $API_KEY)"

360 }

361 }

362 }

363}

364```

365

366### Hooks

367

368Crush has preliminary support for hooks. For details, see

369[the hook guide](./docs/hooks/).

370

371### Ignoring Files

372

373Crush respects `.gitignore` files by default, but you can also create a

374`.crushignore` file to specify additional files and directories that Crush

375should ignore. This is useful for excluding files that you want in version

376control but don't want Crush to consider when providing context.

377

378The `.crushignore` file uses the same syntax as `.gitignore` and can be placed

379in the root of your project or in subdirectories.

380

381### Allowing Tools

382

383By default, Crush will ask you for permission before running tool calls. If

384you'd like, you can allow tools to be executed without prompting you for

385permissions. Use this with care.

386

387```json

388{

389 "$schema": "https://charm.land/crush.json",

390 "permissions": {

391 "allowed_tools": [

392 "view",

393 "ls",

394 "grep",

395 "edit",

396 "mcp_context7_get-library-doc"

397 ]

398 }

399}

400```

401

402You can also skip all permission prompts entirely by running Crush with the

403`--yolo` flag. Be very, very careful with this feature.

404

405### Disabling Built-In Tools

406

407If you'd like to prevent Crush from using certain built-in tools entirely, you

408can disable them via the `options.disabled_tools` list. Disabled tools are

409completely hidden from the agent.

410

411```json

412{

413 "$schema": "https://charm.land/crush.json",

414 "options": {

415 "disabled_tools": ["bash", "sourcegraph"]

416 }

417}

418```

419

420To disable tools from MCP servers, see the [MCP config section](#mcps).

421

422### Disabling Skills

423

424If you'd like to prevent Crush from using certain skills entirely, you can

425disable them via the `options.disabled_skills` list. Disabled skills are hidden

426from the agent, including builtin skills and skills discovered from disk.

427

428```json

429{

430 "$schema": "https://charm.land/crush.json",

431 "options": {

432 "disabled_skills": ["crush-config"]

433 }

434}

435```

436

437### Agent Skills

438

439Crush supports the [Agent Skills](https://agentskills.io) open standard for

440extending agent capabilities with reusable skill packages. Skills are folders

441containing a `SKILL.md` file with instructions that Crush can discover and

442activate on demand.

443

444The global paths we looks for skills are:

445

446* `$CRUSH_SKILLS_DIR`

447* `$XDG_CONFIG_HOME/agents/skills` or `~/.config/agents/skills/`

448* `$XDG_CONFIG_HOME/crush/skills` or `~/.config/crush/skills/`

449* `~/.agents/skills/`

450* `~/.claude/skills/`

451* On Windows, we _also_ look at

452 * `%LOCALAPPDATA%\agents\skills\` or `%USERPROFILE%\AppData\Local\agents\skills\`

453 * `%LOCALAPPDATA%\crush\skills\` or `%USERPROFILE%\AppData\Local\crush\skills\`

454* Additional paths configured via `options.skills_paths`

455

456On top of that, we _also_ load skills in your project from the following

457relative paths:

458

459* `.agents/skills`

460* `.crush/skills`

461* `.claude/skills`

462* `.cursor/skills`

463

464```jsonc

465{

466 "$schema": "https://charm.land/crush.json",

467 "options": {

468 "skills_paths": [

469 "~/.config/crush/skills", // Windows: "%LOCALAPPDATA%\\crush\\skills",

470 "./project-skills",

471 ],

472 },

473}

474```

475

476You can get started with example skills from [anthropics/skills](https://github.com/anthropics/skills):

477

478```bash

479# Unix

480mkdir -p ~/.config/crush/skills

481cd ~/.config/crush/skills

482git clone https://github.com/anthropics/skills.git _temp

483mv _temp/skills/* . && rm -rf _temp

484```

485

486```powershell

487# Windows (PowerShell)

488mkdir -Force "$env:LOCALAPPDATA\crush\skills"

489cd "$env:LOCALAPPDATA\crush\skills"

490git clone https://github.com/anthropics/skills.git _temp

491mv _temp/skills/* . ; rm -r -force _temp

492```

493

494#### User-Invocable Skills

495

496Skills can be made invocable as commands from the commands palette (Ctrl+P). Add `user-invocable: true` to the skill's YAML frontmatter:

497

498```yaml

499---

500name: my-skill

501description: A skill that can be invoked as a command.

502user-invocable: true

503---

504```

505

506User-invocable skills appear in the commands palette with a `user:` or `project:` prefix:

507- Skills from global directories show as `user:skill-name`

508- Skills from project directories show as `project:skill-name`

509

510When invoked, the skill's instructions are loaded into the conversation context.

511

512To prevent the model from auto-triggering a skill (while still allowing user invocation), add `disable-model-invocation: true`:

513

514```yaml

515---

516name: my-skill

517description: Only invocable by users, not the model.

518user-invocable: true

519disable-model-invocation: true

520---

521```

522

523Skills with `disable-model-invocation` won't appear in the model's available skills list but can still be invoked manually by users.

524

525### Desktop notifications

526

527Crush sends desktop notifications when a tool call requires permission and when

528the agent finishes its turn. They're only sent when the terminal window isn't

529focused _and_ your terminal supports reporting the focus state.

530

531```jsonc

532{

533 "$schema": "https://charm.land/crush.json",

534 "options": {

535 "disable_notifications": false, // default

536 },

537}

538```

539

540To disable desktop notifications, set `disable_notifications` to `true` in your

541configuration. On macOS, notifications currently lack icons due to platform

542limitations.

543

544### Initialization

545

546When you initialize a project, Crush analyzes your codebase and creates

547a context file that helps it work more effectively in future sessions.

548By default, this file is named `AGENTS.md`, but you can customize the

549name and location with the `initialize_as` option:

550

551```json

552{

553 "$schema": "https://charm.land/crush.json",

554 "options": {

555 "initialize_as": "AGENTS.md"

556 }

557}

558```

559

560This is useful if you prefer a different naming convention or want to

561place the file in a specific directory (e.g., `CRUSH.md` or

562`docs/LLMs.md`). Crush will fill the file with project-specific context

563like build commands, code patterns, and conventions it discovered during

564initialization.

565

566### Attribution Settings

567

568By default, Crush adds attribution information to Git commits and pull requests

569it creates. You can customize this behavior with the `attribution` option:

570

571```json

572{

573 "$schema": "https://charm.land/crush.json",

574 "options": {

575 "attribution": {

576 "trailer_style": "co-authored-by",

577 "generated_with": true

578 }

579 }

580}

581```

582

583- `trailer_style`: Controls the attribution trailer added to commit messages

584 (default: `assisted-by`)

585 - `assisted-by`: Adds `Assisted-by: Crush:[ModelID]` as specified in [the convention](https://docs.kernel.org/process/coding-assistants.html#attribution)

586 - `co-authored-by`: Adds `Co-Authored-By: Crush <crush@charm.land>`

587 - `none`: No attribution trailer

588- `generated_with`: When true (default), adds `💘 Generated with Crush` line to

589 commit messages and PR descriptions

590

591### Custom Providers

592

593Crush supports custom provider configurations for both OpenAI-compatible and

594Anthropic-compatible APIs.

595

596> [!NOTE]

597> Note that we support two "types" for OpenAI. Make sure to choose the right one

598> to ensure the best experience!

599>

600> - `openai` should be used when proxying or routing requests through OpenAI.

601> - `openai-compat` should be used when using non-OpenAI providers that have OpenAI-compatible APIs.

602

603#### OpenAI-Compatible APIs

604

605Here’s an example configuration for Deepseek, which uses an OpenAI-compatible

606API. Don't forget to set `DEEPSEEK_API_KEY` in your environment.

607

608```json

609{

610 "$schema": "https://charm.land/crush.json",

611 "providers": {

612 "deepseek": {

613 "type": "openai-compat",

614 "base_url": "https://api.deepseek.com/v1",

615 "api_key": "$DEEPSEEK_API_KEY",

616 "models": [

617 {

618 "id": "deepseek-chat",

619 "name": "Deepseek V3",

620 "cost_per_1m_in": 0.27,

621 "cost_per_1m_out": 1.1,

622 "cost_per_1m_in_cached": 0.07,

623 "cost_per_1m_out_cached": 1.1,

624 "context_window": 64000,

625 "default_max_tokens": 5000

626 }

627 ]

628 }

629 }

630}

631```

632

633#### Anthropic-Compatible APIs

634

635Custom Anthropic-compatible providers follow this format:

636

637```json

638{

639 "$schema": "https://charm.land/crush.json",

640 "providers": {

641 "custom-anthropic": {

642 "type": "anthropic",

643 "base_url": "https://api.anthropic.com/v1",

644 "api_key": "$ANTHROPIC_API_KEY",

645 "extra_headers": {

646 "anthropic-version": "2023-06-01"

647 },

648 "models": [

649 {

650 "id": "claude-sonnet-4-20250514",

651 "name": "Claude Sonnet 4",

652 "cost_per_1m_in": 3,

653 "cost_per_1m_out": 15,

654 "cost_per_1m_in_cached": 3.75,

655 "cost_per_1m_out_cached": 0.3,

656 "context_window": 200000,

657 "default_max_tokens": 50000,

658 "can_reason": true,

659 "supports_attachments": true

660 }

661 ]

662 }

663 }

664}

665```

666

667### Amazon Bedrock

668

669Crush currently supports running Anthropic models through Bedrock, with caching disabled.

670

671- A Bedrock provider will appear once you have AWS configured, i.e. `aws configure`

672- Crush also expects the `AWS_REGION` or `AWS_DEFAULT_REGION` to be set

673- To use a specific AWS profile set `AWS_PROFILE` in your environment, i.e. `AWS_PROFILE=myprofile crush`

674- Alternatively to `aws configure`, you can also just set `AWS_BEARER_TOKEN_BEDROCK`

675

676### Vertex AI Platform

677

678Vertex AI will appear in the list of available providers when `VERTEXAI_PROJECT` and `VERTEXAI_LOCATION` are set. You will also need to be authenticated:

679

680```bash

681gcloud auth application-default login

682```

683

684To add specific models to the configuration, configure as such:

685

686```json

687{

688 "$schema": "https://charm.land/crush.json",

689 "providers": {

690 "vertexai": {