Default Settings for Kimi-CLI: A Complete Guide to Configuration Values

The default settings for kimi-cli include a web server port of 5494, 50,000-character tool output limits, 32,000 fallback tokens for LLM context budgeting, and telemetry enabled by default, all of which can be overridden via environment variables or TOML configuration files.

The MoonshotAI/kimi-cli repository defines its out-of-the-box behavior through a centralized configuration system located in src/kimi_cli/config.py. These defaults govern everything from logging verbosity to token safety margins, ensuring the CLI operates securely and efficiently without requiring immediate user configuration.

Core Configuration File Locations

Before examining specific values, it is essential to understand where kimi-cli expects to find its settings. According to src/kimi_cli/config.py, the tool uses the XDG Base Directory specification with a fallback to the home directory.

# From src/kimi_cli/config.py

DEFAULT_CONFIG_DIR: Final[Path] = Path(os.getenv("XDG_CONFIG_HOME", Path.home() / ".config")) / "kimi"
DEFAULT_CONFIG_FILE: Final[Path] = DEFAULT_CONFIG_DIR / "config.toml"
DEFAULT_SECRETS_FILE: Final[Path] = DEFAULT_CONFIG_DIR / "secrets.toml"
ENV_KIMI_CONFIG_DIR: Final[str] = "KIMI_CONFIG_DIR"

The KIMI_CONFIG_DIR environment variable overrides the default directory, allowing users to relocate their configuration to custom paths.

General Runtime Defaults

The Config class in src/kimi_cli/config.py defines the primary runtime behavior using Pydantic models with sensible fallbacks.

Logging and Telemetry

By default, kimi-cli operates with transparency and basic analytics:

• log_level: Set to "INFO" by default, controlling console verbosity for debugging and operational feedback.
• telemetry: Enabled (True) by default, collecting anonymous usage data to improve the tool.

Web Interface Port

The default port for the FastAPI-based web server is defined consistently across the codebase:

• default_port: 5494 (defined in src/kimi_cli/config.py and mirrored in src/kimi_cli/web/app.py as DEFAULT_PORT)

A separate visualization server uses port 5495 as defined in src/kimi_cli/vis/app.py.

LLM Token Budgeting Defaults

Token management is critical for preventing context window overflows. The defaults in src/kimi_cli/config.py and src/kimi_cli/llm.py provide conservative safety margins:

# From src/kimi_cli/config.py

unknown_context_completion_tokens: int = Field(default=32_000)
completion_token_safety_margin: int = Field(default=1_024)

• unknown_context_completion_tokens: 32_000 tokens serve as the fallback budget when the underlying LLM provider does not report its context size.
• completion_token_safety_margin: 1_024 tokens are reserved as a buffer, subtracted from the model's maximum token limit to prevent truncation errors.

These values are duplicated in src/kimi_cli/llm.py as DEFAULT_UNKNOWN_CONTEXT_COMPLETION_TOKENS and DEFAULT_COMPLETION_TOKEN_SAFETY_MARGIN, and enforced within the core agent loop in src/kimi_cli/soul/kimisoul.py.

Execution Flow Limits

To prevent runaway agent behavior, kimi-cli implements hard limits on execution steps:

• max_flow_moves: 1_000 (defined in src/kimi_cli/config.py and src/kimi_cli/soul/kimisoul.py as DEFAULT_MAX_FLOW_MOVES)

This setting caps the number of tool invocations within a single agent flow, protecting against infinite loops.

Security and CORS Defaults

The web server implements strict security defaults in src/kimi_cli/web/auth.py:

DEFAULT_ALLOWED_ORIGIN_REGEX: Final[re.Pattern] = re.compile(
    r"^https?://(localhost|127\.0\.1)(:\d+)?$"
)

This regular expression restricts Cross-Origin Resource Sharing (CORS) to localhost origins only (http://localhost, http://127.0.0.1), mitigating CSRF attacks when running the server locally.

Session API Path Depth

The session management API enforces directory traversal protection:

• DEFAULT_MAX_PUBLIC_PATH_DEPTH: 6 (from src/kimi_cli/web/api/sessions.py)

This limits how many path components can be exposed through the public API, preventing access to deeply nested system directories.

Tool Output Constraints

Tool outputs are truncated to prevent console flooding and token overflow, as defined in src/kimi_cli/tools/utils.py:

• DEFAULT_MAX_CHARS: 50_000 characters (global output limit)
• DEFAULT_MAX_LINE_LENGTH: 2_000 characters (per-line limit)

The truncate_output() function applies the line-length limit first, then enforces the global character cap, appending an ellipsis (…) when truncation occurs.

Agent Specification Defaults

Kimi-cli uses YAML-based agent specifications with built-in fallbacks in src/kimi_cli/agentspec.py:

• DEFAULT_AGENT_SPEC_VERSION: "1"
• DEFAULT_AGENT_FILE: src/kimi_cli/agents/default/agent.yaml
• OKABE_AGENT_FILE: src/kimi_cli/agents/okabe/agent.yaml (legacy compatibility)

When users do not specify a custom --agent-file, the system loads the default agent specification from the package directory.

LLM Provider Configuration

The abstraction layer in src/kimi_cli/llm.py defaults to OpenAI-compatible endpoints:

• Provider: "openai"
• Model: "gpt-4o-mini"
• Environment override: KIMI_LLM_PROVIDER can change the default provider without modifying configuration files.

Overriding Default Settings

Users can customize these defaults through multiple mechanisms. The configuration loader in src/kimi_cli/config.py uses the following precedence:

1. Environment variables (e.g., KIMI_CONFIG_DIR, KIMI_LLM_PROVIDER)
2. Configuration file (~/.config/kimi/config.toml or $KIMI_CONFIG_DIR/config.toml)
3. Hard-coded defaults (as detailed above)

Example config.toml to override key defaults:

log_level = "DEBUG"
telemetry = false
default_port = 8080
max_flow_moves = 500
unknown_context_completion_tokens = 16_000

Summary

• Configuration files default to ~/.config/kimi/ (or $XDG_CONFIG_HOME/kimi/), with KIMI_CONFIG_DIR available for relocation.
• Web server defaults to port 5494 with strict localhost-only CORS policies and a maximum public path depth of 6.
• Token budgeting uses a 32,000 token fallback budget and a 1,024 token safety margin to prevent context overflow.
• Tool outputs are limited to 50,000 characters total and 2,000 characters per line.
• Agent flows are capped at 1,000 moves to prevent infinite execution loops.
• LLM defaults target OpenAI's gpt-4o-mini but can be overridden via KIMI_LLM_PROVIDER.

Frequently Asked Questions

How do I change the default port for the kimi-cli web server?

The default port is 5494, defined in src/kimi_cli/config.py as default_port and in src/kimi_cli/web/app.py as DEFAULT_PORT. You can override this by setting default_port = 8080 in your config.toml file, or by passing the --port flag when running kimi server.

Where does kimi-cli store its configuration files by default?

According to src/kimi_cli/config.py, kimi-cli stores configuration in DEFAULT_CONFIG_DIR, which resolves to $XDG_CONFIG_HOME/kimi/ or ~/.config/kimi/ on most systems. The main configuration file is config.toml, and sensitive values belong in secrets.toml. You can relocate these by setting the KIMI_CONFIG_DIR environment variable.

What are the default token limits for LLM interactions in kimi-cli?

The defaults are defined in both src/kimi_cli/config.py and src/kimi_cli/llm.py. The unknown_context_completion_tokens default is 32,000, used when the model does not report its context size. Additionally, completion_token_safety_margin defaults to 1,024 tokens, which is subtracted from the model's maximum capacity to create a safety buffer against overflow.

Is telemetry enabled by default in kimi-cli?

Yes. The telemetry field in the Config class (src/kimi_cli/config.py) defaults to True, enabling anonymous usage collection. You can disable this by setting telemetry = false in your config.toml file.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}