Skip to main content
Kilo Code is a coding agent for editors such as VS Code and Cursor, with support for multiple model providers, code changes, debugging, and terminal tasks. Its Step Plan provider connects these workflows to Step models.

​
Prerequisites

​
Install Kilo Code

Find Kilo Code in your editor’s extension marketplace and install it. The extension ID is kilocode.kilo-code. These steps use a version with Providers and Models settings pages.

​
Configure Step Plan

​
Connect the built-in provider

  1. Open Kilo Code settings, select Providers, then choose StepFun Step Plan (Global).
  2. Enter your API key and click Submit. Do not include a Bearer prefix or leading or trailing spaces in the API Key field.
  3. Under Settings > Models > Default Model, select Step 5 Preview from that provider. Its model ID is step-5-preview.
  4. Save and start a new task. Keep the built-in provider’s model presets.

​
Use a custom provider

If the Step Plan preset or your target model is missing, open Providers > Custom provider and enter: Click Submit, then select the model under your custom provider. This configuration uses Chat Completions at /step_plan/v1/chat/completions. Manually added models also need context and output capacity information. If these fields are missing, merge the following into kilo.json or kilo.jsonc, preserving the provider’s existing connection and authentication settings:
These capacities apply to Step 5 Preview. Match them to the model when switching. Reasoning models need room for both reasoning and visible output; a small output budget may produce empty or truncated text.

​
Connect StepSearch MCP

StepSearch provides web_search and web_fetch through the Step Plan endpoint for web searches and page retrieval.

​
Set the authentication variable

Fully quit the editor, then set the key and start the editor from your project’s terminal. Replace YOUR_STEP_API_KEY with your key. For Cursor, replace code . with cursor ..
If the editor command is unavailable, configure its command-line launcher first. The environment variable applies to this terminal and processes it starts. Make sure the editor can read STEP_API_KEY whenever you use this MCP server.

​
Add the MCP server

  1. Open MCP Servers in Kilo Code settings.
  2. Edit kilo.json in the current project. If the project already uses kilo.jsonc or one of these files under .kilo/, merge into that file instead, preserving other settings:
  1. Save and reconnect the server. Confirm that step-search is connected and lists web_search and web_fetch.
  2. Start a new task using either tool, approve the call when prompted, and inspect the returned result.
The current format uses the top-level mcp key, type: "remote", and {env:STEP_API_KEY} for environment-variable substitution. Remote connections try Streamable HTTP first and fall back to SSE if needed. This example authenticates through a request header, not OAuth.
Keep the environment-variable reference in the project configuration. Do not replace it with a real key and commit the file. Anyone with the key may make requests using your permissions and consume quota.

​
Verify the integration

​
Test a conversation

Start a new task and enter Reply only with OK. Expected output: OK.

​
Test tool calls

  1. Create a separate test directory with a README.md containing Kilo Code Step Plan test, then open it in the editor.
  2. Start a new task and enter:
  1. Approve the required file operations and commands. Inspect the tool records, confirm that hello.py exists and prints Hello, world!, and check that README.md is unchanged.

​
Check saved settings

Restart the editor, confirm that the provider and model selection are retained, and start another short conversation.

​
Common questions

Update Kilo Code and check the provider list. Alternatively, open Providers > Custom provider, select OpenAI Compatible, enter the Step Plan base URL, and add step-5-preview. Save and select the model under that custom provider.
Inspect the full error, network connection, base URL, API key, and model permissions. Custom providers use https://api.stepfun.ai/step_plan/v1. The key must be complete, valid, and associated with the account for this endpoint. Use the server response to distinguish authentication, model, and request-parameter failures.
Check the stop reason and output budget. Reasoning may consume the budget before visible text is produced. Restore presets for a built-in model, or check capacity and request settings for a custom model, then adjust based on the actual error.
Include max output tokens and Max Output Tokens = -1 belong to an older interface, not the current built-in provider setup. Current custom models describe capacity with limit.context and limit.output. Use valid capacities for your model rather than copying -1 into limit.output.
The current setup uses mcp in kilo.json or kilo.jsonc. Migrate older mcpServers, streamable-http, and disabled fields to the current format, or keep a configuration matched to the older extension version. Do not mix the two schemas.
Check that you edited the configuration file loaded by the current project, with type: "remote" and enabled: true. Confirm that the editor process can read STEP_API_KEY and the header is Bearer {env:STEP_API_KEY}. Save, reconnect, and inspect the server log.
Open the target project, select a mode that allows file changes, and check for pending tool approvals. Approve the required operations; enabling every auto-approval option is unnecessary.