This project is a fork of vim-ai by madox2. I would like to express my gratitude to the original author for their work.

• Replaced Python code with TypeScript.
• Integrated with coc.nvim for async AI features.

This project uses yarn for package management during development.

1. Ensure your vim is newer than 9 and node is newer than 18: node --version, hint for miserable CentOS users: Node.js unofficial-builds project
2. Install coc.nvim. Checkout Install coc.nvim for more infomation.
3. Use command :CocInstall https://github.com/Kuro96/coc-ai in your vim

See roles-example.toml. Make a copy of your own and specify path to it with coc-ai.global.rolesConfigPath.

Available Options, all options are optional:

[<your role name>]
prompt = "<extra prompt which will be add to user prompt>"

[<your role name>.options]
model = "<model name>"
endpointUrl = "<protocol>://<hostname>/v1/chat/completions"
requiresAuth = true # <true/false>
authProvider = "api-key" # "api-key" or "openai-codex"
tokenPath = "<path to your token file>"
oauthTokenPath = "~/.opencode/auth/openai.json" # for authProvider = "openai-codex"; falls back to ~/.local/share/opencode/auth.json, then ~/.codex/auth.json when unset/default path is missing
proxy = "<protocol>://<ip>:<port>"
maxTokens = 4096  # <max token num>
temperature = 0.6  # <temperature>
thinking = { type = "enabled" } # OpenAI-compatible thinking switch: "enabled" or "disabled"
reasoningEffort = "high" # sent as reasoning_effort; native values are preserved unless a provider needs compatibility mapping
initialPrompt = "<overrides `initialPrompt` in coc-settings>"

[<your role name>.options-chat]
# same as options, overrides for `AIChat` command

[<your role name>.options-complete]
# same as options, overrides for `AI` command

[<your role name>.options-edit]
# same as options, overrides for `AIEdit` command

Same as role options, but ONLY available with AIChat command.

Put it on top of ai chat window, or simply modify after one conversation with coc-ai.chat.populatesOptions set to true.

coc-settings.json, see docs of coc.nvim for more info. You can edit it by command :CocConfig.

• coc-ai.tab.enabled: boolean (default: false). Enable/disable AI tab completion.
• coc-ai.chat.autoTitle: boolean (default: true). Automatically title the chat buffer after the first exchange.
• coc-ai.*.thinking: object (example: { "type": "enabled" }). OpenAI-compatible thinking mode switch for providers that support it; omitted for OpenAI GPT requests.
• coc-ai.*.reasoningEffort: string (example: "minimal", "low", "high", or "xhigh"). Sent as OpenAI-compatible reasoning_effort; provider compatibility maps unsupported aliases when needed.
• coc-ai.*.authProvider: string ("api-key" or "openai-codex"). "openai-codex" reuses an OpenCode/Codex OAuth token to access ChatGPT Plus/Pro subscription models through the Codex backend.
• coc-ai.*.oauthTokenPath: string (default: "~/.opencode/auth/openai.json"). Token file created by OpenCode Codex auth plugins; when the default path is missing, coc-ai falls back to modern OpenCode auth at "~/.local/share/opencode/auth.json", then the official Codex CLI cache at "~/.codex/auth.json". See roles-example.toml for GPT 5.5 / 5.4 mini presets that also demonstrate proxy.

For ChatGPT Plus/Pro subscription access, sign in once with OpenCode/Codex first so an OAuth cache exists, then run :AIConfig and choose Provider/model wizard -> ChatGPT Plus/Pro (Codex OAuth) to select the concrete model for global, chat, edit, complete, or tab. Supported file caches are OpenCode plugin auth at ~/.opencode/auth/openai.json, modern OpenCode auth at ~/.local/share/opencode/auth.json, and official Codex CLI auth at ~/.codex/auth.json.

:AIConfig is a small configuration center inspired by OpenCode's discoverable config workflow, adapted to coc.nvim's Vim UI constraints. Coc does not provide a full multi-pane settings TUI inside Vim, so coc-ai keeps common changes in menu/input prompts and sends complex edits back to normal files:

• Provider/model wizard: configure ChatGPT Plus/Pro OAuth, OpenAI-compatible API keys, or just the model for global, chat, edit, complete, or tab.
• Edit one common setting: update frequently changed values such as model, endpoint, proxy, token paths, reasoning effort, chat UI behavior, and tab completion without remembering JSON keys.
• Set roles config path: change coc-ai.global.rolesConfigPath interactively.
• Open roles TOML: open the roles file in Vim, creating a starter file if it does not exist.
• Open coc-settings.json: jump to :CocConfig for advanced or bulk edits.

The /chatgpt-plan role is intentionally only an auth/backend mixin for ChatGPT subscription plans. It does not pick a model by itself; configure coc-ai.*.model with :AIConfig, :CocConfig, or another role, then use /chatgpt-plan when you want that request to go through the Codex OAuth backend.

Also you can refer to my example below:

{
  "coc-ai.global.model": "deepseek-v4-pro",
  "coc-ai.global.endpointUrl": "https://api.deepseek.com/v1/chat/completions",
  "coc-ai.global.thinking": { "type": "enabled" },
  "coc-ai.global.reasoningEffort": "high",
  "coc-ai.global.requiresAuth": true,
  "coc-ai.global.tokenPath": "~/.vim/token",
  "coc-ai.global.rolesConfigPath": "~/.vim/vimrc.d/coc-ai-roles.toml",
  "coc-ai.chat.autoScroll": true,
  "coc-ai.chat.scratchBufferKeepOpen": false,
  "coc-ai.chat.populatesOptions": true,
  "coc-ai.chat.preserveFocus": true,
  "coc-ai.chat.openChatCommand": "preset_right",
  "coc-ai.chat.autoTitle": true,
  "coc-ai.tab.enabled": true
}

role options > chat options > Coc options

1. AIChat: :[%]['<,'>]AIC[hat] [/<role>] [<prompt>]
2. AIEdit: :[%]['<,'>]AIE[dit] [/<role>] [<prompt>]
3. AIComplete: :[%]['<,'>]AI [/<role>] [<prompt>]
4. AIStop: :AIS[top]
5. AIBack(resume hidden chat window if scratchBufferKeepOpen set to true): :AIB[ack]
6. AIConfig: :AIConfig opens the interactive configuration center for provider/model/auth setup, common settings, roles TOML, and coc-settings.json.

NOTE:

• any characters inside [] is optional
• :% equal to :1,$, checkout :help :%, as well as :help '< :help '>
• /role MUST precede prompt
• /role can be specified multiple times, and the latter configuration will override the former.

:AIC hello world
:AIC /v4pro who are you?
:%AIC /gpt55 /explain 用中文回答
:%AIC /gpt54mini /explain 用中文回答
:AIConfig
:AIC /chatgpt-plan implement this with my configured ChatGPT subscription model

• Proxy for api request
• Option for chat whether keep track with cursor
• Implement ai complete
• Implement ai edit
• Support CocInstall
• README
• Auto attach with .aichat files
• Implemented AI tab completion
• Added optional auto-titling for AIChat.

MIT

This extension is built with create-coc-extension