Kimi for Copilot Chat
Use Kimi models from the normal Copilot Chat model picker.
This extension is a BYOK provider. It does not add a second chat sidebar or a local proxy. Pick a Kimi model in Copilot Chat and keep using Copilot's agent mode, tools, instructions, and MCP setup.
Repository: https://github.com/lucatotem/Kimi-Extension
What It Adds
- Kimi models in the Copilot Chat model picker
- API key storage through VS Code SecretStorage
- Optional one-time import from
KIMI_API_KEY or MOONSHOT_API_KEY
- A switchable API mode for Kimi Platform, Kimi Code Plan, or custom proxies
- Complete endpoint-specific model catalogs for Kimi Platform and Kimi Code
- Native Copilot thinking blocks for Kimi
reasoning_content
- Tool calls and tool results for agent mode
- Image and video content for Kimi models that support multimodal input
- Optional model discovery through Kimi's
/v1/models endpoint
- Model ID overrides for compatible proxies
Models
The picker follows the catalog of the active endpoint. Equivalent models use different API IDs on the two services, so the extension keeps the catalogs separate and sends the native ID shown below.
Catalog verified on October 26, 2026 against the official Platform model list, parameter reference, and Kimi Code model configuration.
| Model |
API ID |
Use for |
| Kimi K3.1 Pro |
kimi-k3.1-pro |
Latest flagship with vision, tools, adjustable reasoning, and 1M context. |
| Kimi K3.1 Pro Highspeed |
kimi-k3.1-pro-highspeed |
The same capabilities with faster output. |
| Kimi K3 |
kimi-k3 |
Previous flagship with vision, adjustable reasoning, and 1M context. |
| Kimi K2.7 Code |
kimi-k2.7-code |
Coding and agent tasks with thinking always on. |
| Kimi K2.7 Code High-Speed |
kimi-k2.7-code-highspeed |
The K2.7 coding model at higher output speed. |
| Kimi K2.6 |
kimi-k2.6 |
General coding, agent, text, image, and video tasks with optional thinking. |
Kimi K2.5 and every Moonshot V1 variant were retired on August 31, 2026 and are excluded from the bundled picker and official discovery results. Older K2 preview/thinking models, kimi-latest, and kimi-thinking-preview are also excluded.
Platform K2.6 and K2.7 Code models are still available, so they remain selectable and are labeled with their planned December 2026 retirement. Kimi says these models will stop being enabled for newly registered users on October 31, 2026 and are planned to retire by the end of December. Use K3.1 Pro for new integrations.
Kimi Code Plan
| Model |
API ID |
Availability |
| Kimi K3.1 Pro |
k3.1-pro |
Plus / Moderato and above; 1M context requires Pro / Allegretto and above. |
| Kimi K3.1 Pro Highspeed |
k3.1-pro-highspeed |
Pro / Allegretto and above; up to 1M context. |
| Kimi K3 |
k3 |
Plus / Moderato and above; 1M context requires Pro / Allegretto and above. |
| Kimi K3 256K |
k3-256k |
Plus / Moderato and above; fixed 256K context and no video input. |
| Kimi K2.8 Preview |
kimi-for-coding |
Plus and above on new plans, Andante and above on legacy plans; up to 1M context. |
| Kimi K2.7 Code High-Speed |
kimi-for-coding-highspeed |
Pro / Allegretto and above; 256K context. |
kimi-for-coding now serves K2.8 Preview, replacing K2.7 Code without changing the API ID. It supports low, high, and max reasoning effort, with max as the default. Go plans do not include coding quota.
Kimi Code does not expose K2.6, K2.5, or Moonshot V1 as selectable API models. Kimi Platform does not expose the kimi-for-coding aliases. The extension therefore shows every documented model available on the active service rather than offering entries that the endpoint will reject.
When an API key is configured, the extension also merges models returned by the active endpoint's GET /v1/models response. This allows newly released models to appear without an extension update. Discovery updates account-specific context limits when provided; bundled limits represent the maximum supported context and do not grant plan access. Custom proxies can still expose older models through discovery if they host them themselves.
API Modes
Use Kimi: Switch API Mode or the kimi-copilot.apiMode setting to choose the backend:
| Mode |
Endpoint |
Key command |
Notes |
| Kimi Platform (Pay as you go) |
https://api.moonshot.ai/v1 |
Kimi: Set API Key |
Default mode for existing pay-as-you-go API keys. |
| Kimi Code Plan |
https://api.kimi.com/coding/v1 |
Kimi: Set Kimi Code API Key |
Offers the six documented Kimi Code model IDs and uses a separate subscription key. |
| Custom Base URL |
kimi-copilot.baseUrl |
Kimi: Set API Key |
For OpenAI-compatible proxies or manually configured endpoints. |
Switching modes does not delete stored keys. The regular pay-as-you-go key is stored separately from the Kimi Code key, so users can move between modes without re-entering the other key.
To use the overseas Kimi Code endpoint, choose Custom Base URL and set kimi-copilot.baseUrl to https://api.kimi.ai/coding/v1. Both official Code domains use the Kimi Code catalog and its separate key. A URL ending in /coding is normalized to /coding/v1.
To use a Kimi Code subscription plan:
- Run
Kimi: Switch API Mode.
- Choose
Kimi Code Plan.
- Run
Kimi: Set Kimi Code API Key.
- Run
Kimi: Refresh Models or reload VS Code.
- Pick K3.1 Pro, K3.1 Pro Highspeed, K3, K3 256K, K2.8 Preview, or K2.7 Code High-Speed according to your membership tier.
To use the pay-as-you-go Kimi Platform:
- Run
Kimi: Switch API Mode.
- Choose
Kimi Platform (Pay as you go).
- Run
Kimi: Set API Key.
- Run
Kimi: Refresh Models or reload VS Code.
- Pick a model, for example
Kimi K3.1 Pro, from the Copilot Chat model picker.
Using both plans together
The extension stores the platform key and the Kimi Code key separately, so both plans can stay configured at the same time:
- Run
Kimi: Set API Key once for your platform key and Kimi: Set Kimi Code API Key once for your Kimi Code plan key.
- Switch backends at any time with
Kimi: Switch API Mode. Switching never deletes either key.
- Each mode shows the complete catalog supported by its endpoint. Shared model names are sent with that endpoint's native ID, such as
kimi-k3 on Platform and k3 on Kimi Code.
- Use
kimi-copilot.modelIdOverrides only when a compatible proxy or a new API alias requires a different ID.
- Model discovery runs against the active endpoint whenever its key is set, so newly available models are merged into that mode's picker.
For diagnostics, run Kimi: Test Connection and then Kimi: Show Logs. The test reports whether the endpoint returned an HTTP status, or whether the request failed before reaching HTTP.
Thinking
Kimi thinking models stream reasoning_content before the final answer. The extension reports those chunks as Copilot thinking parts, so they render separately from the response.
The Copilot model picker shows a Thinking control where Kimi supports it:
| Value |
Kimi request |
| None |
thinking: { "type": "disabled" } on Platform K2.6 and supported Code models. Code K3 / K3 256K / K2.7 High-Speed route to K2.8 Preview without thinking. |
| Low |
reasoning_effort: "low" on K3.1 Pro, K3, and Code K2.8 Preview. |
| High |
reasoning_effort: "high" where supported; otherwise enables thinking where configurable. |
| Max |
reasoning_effort: "max" where supported; otherwise enables preserved thinking where configurable. |
K3.1 Pro always thinks on both endpoints. Platform K3 and Platform K2.7 Code also always think. The picker never offers None for these models. K3.1 Pro and K3 default to max on Platform and high on Code; Code K2.8 Preview defaults to max. Code K2.7 High-Speed does not support adjustable reasoning effort.
Start a new chat after switching models or reasoning effort, especially when moving to K3.1 Pro, as recommended by Kimi's documentation.
Commands
| Command |
What it does |
Kimi: Switch API Mode |
Choose Kimi Platform, Kimi Code Plan, or Custom Base URL |
Kimi: Set API Key |
Store your Kimi Platform or custom endpoint API key |
Kimi: Set Kimi Code API Key |
Store your Kimi Code subscription API key |
Kimi: Open API Keys |
Open the Kimi API key page |
Kimi: Clear API Key |
Remove the stored Kimi Platform/custom key |
Kimi: Clear Kimi Code API Key |
Remove only the stored Kimi Code key |
Kimi: Open Settings |
Open extension settings |
Kimi: Show Logs |
Open the Kimi output channel |
Kimi: Test Connection |
Test /models and write the HTTP result or fetch failure cause to logs |
Kimi: Open Request Dumps Folder |
Open verbose request dumps |
Kimi: Refresh Models |
Clear the model cache and run discovery again |
Settings
| Setting |
Default |
Notes |
kimi-copilot.apiMode |
platform |
Selects Kimi Platform, Kimi Code Plan, or Custom Base URL. |
kimi-copilot.baseUrl |
https://api.moonshot.ai/v1 |
Used when apiMode is custom. Older installs with a custom baseUrl and no apiMode setting keep using that URL. |
kimi-copilot.maxTokens |
0 |
0 leaves the output limit to the API default. |
kimi-copilot.enableModelDiscovery |
true |
Adds models returned by the active endpoint's /v1/models response when its key is set. |
kimi-copilot.modelIdOverrides |
official IDs |
Maps native picker IDs to alternate API IDs, primarily for compatible proxies. |
kimi-copilot.debugMode |
minimal |
verbose writes full request dumps. Use it only while diagnosing issues. |
API Key Storage
Use Kimi: Set API Key once for Kimi Platform or custom endpoints. Use Kimi: Set Kimi Code API Key for the Kimi Code plan. Both keys are stored with VS Code SecretStorage, which uses the operating system keychain where VS Code supports it.
You can also start VS Code with one of these environment variables set:
KIMI_API_KEY=sk-...
MOONSHOT_API_KEY=sk-...
KIMI_CODE_API_KEY=sk-...
When the extension sees one of those variables and no matching key is already stored, it imports the value into SecretStorage. KIMI_API_KEY and MOONSHOT_API_KEY import into the regular key slot; KIMI_CODE_API_KEY imports into the Kimi Code key slot. After that, you do not need to recreate or re-enter the key each time you use the extension.
Development
npm install
npm run compile
Press F5 in VS Code to launch an Extension Development Host.
Notes
Kimi rejects empty user messages. This extension drops empty text-only messages and preserves assistant tool calls with empty content, which is the shape expected by OpenAI-compatible tool-calling APIs.
Development note: this extension was built by ivanray with AI assistance. The goal is to match the Copilot Chat BYOK provider experience while keeping the Kimi-specific request handling clear and reviewable.
License
MIT
