Skip to content

kimi.md

Latest commit

 

History

History
105 lines (71 loc) · 4.08 KB

kimi.md

File metadata and controls

105 lines (71 loc) · 4.08 KB
Raw

Kimi Code CLI — Agent Backend Guide

Kimi Code CLI is Moonshot AI's coding agent. It supports the Agent Client Protocol (ACP) natively through kimi acp, so OpenAB can launch it directly over stdio without an adapter.

Prerequisites

  • Kimi Code CLI installed, or the OpenAB Kimi image
  • A Kimi Code account or a Moonshot/Kimi Platform API key
  • Persistent storage for Kimi credentials and configuration in containerized deployments

Configuration

[agent]
command = "kimi"
args = ["acp"]
working_dir = "/home/node"

kimi acp communicates with OpenAB using JSON-RPC over stdin/stdout. Kimi's logs are written to stderr, keeping the ACP stream clean.

Authentication

Kimi Code uses an interactive login flow. In a running container, start the CLI and enter /login:

kubectl exec -it deployment/openab-kimi -- kimi
# Select Kimi Code OAuth or Kimi Platform API key, then follow the prompts.

Credentials and configuration are stored under ~/.kimi-code/. The Kimi image uses /home/node as HOME, so enable persistence for production deployments. For non-container usage, run kimi once in the desired home directory and complete /login before starting OpenAB.

The chart/image exposes the same helper used by other backends:

kubectl exec -it deployment/openab-kimi -- sh -c "$OPENAB_AGENT_AUTH_COMMAND"

Then enter /login in the interactive Kimi session.

Docker

Build the standalone image:

docker build -f Dockerfile.kimi -t openab-kimi .

Build the unified image target:

docker build --target kimi -f Dockerfile.unified -t openab-kimi .

The image pins @moonshot-ai/kimi-code to a specific version. Update KIMI_CODE_VERSION in the Dockerfiles through a dedicated version-bump change.

Helm

Use the Kimi image variant and provide the raw OpenAB configuration through configToml or configUrl:

agents:
  kimi:
    enabled: true
    image: ghcr.io/openabdev/openab:beta-kimi
    workingDir: /home/node
    configToml: |
      [discord]
      allowed_channels = ["YOUR_CHANNEL_ID"]

      [agent]
      command = "kimi"
      args = ["acp"]
      working_dir = "/home/node"

For a complete deployment, inject DISCORD_BOT_TOKEN through a Secret and enable a persistent volume. Discord IDs should be passed with Helm's --set-string option when setting them on the command line.

Model providers

Kimi Code can use its managed Kimi Code OAuth service or a Kimi/Moonshot API key. It also supports configuring Anthropic, OpenAI-compatible, OpenAI Responses, Google Gemini, and Vertex AI providers in ~/.kimi-code/config.toml. See the Kimi providers and models reference.

Example Kimi provider configuration:

[providers.kimi]
type = "kimi"
base_url = "https://api.moonshot.ai/v1"

[providers.kimi.env]
KIMI_API_KEY = "sk-..."

Note that Kimi Code reads provider credentials only from config.toml — the provider's api_key field or its [providers.<name>.env] subtable — and does not fall back to shell environment variables. Supply the key through the agent's configToml (or the Kimi credential store) rather than relying on OpenAB's [agent].inherit_env/[agent].env, which control the broker's child-process environment and are not read by Kimi for provider credentials.

ACP capabilities

Kimi Code's ACP adapter supports normal session creation/loading, prompts, tool calls, permission handling, image input, and MCP forwarding. Audio prompts are not currently supported by Kimi's ACP adapter.

Troubleshooting

  • The process exits immediately: run kimi interactively and complete /login first.
  • kimi cannot be found: use the Kimi image or set [agent].command to the absolute path returned by which kimi.
  • Authentication disappears after restart: enable the agent PVC and verify that HOME=/home/node is writable.
  • No response from ACP: run kimi acp directly and verify that it waits for JSON-RPC input rather than printing an interactive banner.