Skip to content
Mixen API
RU

Kilo Code

Connect Kilo Code to Mixen — pay from your balance in roubles, with a catalog of 105+ models.

Installation

Section titled “Installation”
  • VS Code — open Extensions (Ctrl+Shift+X / Cmd+Shift+X), search for “Kilo Code” and click Install.
  • JetBrains — Settings → Plugins → Marketplace tab, search for “Kilo Code” and install it.

Kilo Code also ships as a CLI; all clients read and write the same shared kilo.jsonc config — a provider added in one is available in the others. Details — in the official documentation.

Configuration

Section titled “Configuration”
  1. Kilo Code panel → Settings (gear icon) → Providers tab.
  2. Scroll the provider list to the bottom and click Custom provider.
  3. Fill in the dialog:
    • Provider ID — mixen
    • Display name — Mixen
    • Provider API — OpenAI Compatible
    • Base URL — https://api.mixen.ai/v1
    • API key — your mxn-… key
    • Models — Kilo will fetch the list from /v1/models automatically: pick models with the (fuzzy) search, or type an ID manually
  4. Click Submit — the selected models appear in the model picker.

If auto-detection doesn’t work, type the model ID manually from the catalog — for example, openai/gpt-5.6-sol.

Key points:

  • OpenAI Compatible is the Chat Completions protocol: Kilo will call /v1/chat/completions.
  • Fine-grained model options (token limits, tool calling, variants) are edited in the kilo.jsonc file.
  • On JetBrains the plugin uses the same shared kilo.jsonc — the provider and models from VS Code apply automatically.

Verifying the connection

Section titled “Verifying the connection”

Before setting up the provider, you can check the key and balance with a minimal generating request:

curl https://api.mixen.ai/v1/chat/completions \
  -H "Authorization: Bearer your-api-key" \
  -H "Content-Type: application/json" \
  -d '{"model":"z-ai/glm-5.3-flash","messages":[{"role":"user","content":"ping"}],"max_tokens":5}'

200 with a response — the key and balance work; 401 invalid api key — the key was copied incompletely or isn’t a Mixen API key; 402 — no funds on the balance.

GET /v1/models does not verify the key — it responds without one; use it only to check that the API is reachable.

Picking a model per task

Section titled “Picking a model per task”

Kilo Code keeps a separate model per agent, so the “cheap for planning, strong for editing” pairing is configured once. Prices are per 1M tokens, input/output; full tables with cache columns — in the catalog.

Task Model Input / output
Reconnaissance: code questions, search, planning GLM 5.3 Flash (z-ai/glm-5.3-flash) 8.4 / 27.9 ₽
Edits in familiar code, routine tasks GLM 5.3 (z-ai/glm-5.3) 156 / 490.2 ₽
Multi-step agentic tasks, refactoring GPT-5.6 Sol (openai/gpt-5.6-sol), Claude Sonnet 5 (anthropic/claude-sonnet-5) 213.9 / 1069.5 ₽

GLM 5.3 Flash costs like a “free” model but holds up in the agent protocol — convenient to leave on Ask and Plan while Code runs a flagship.

Agent Task Model
Plan planning and architectural decisions z-ai/glm-5.3-flash
Ask code questions z-ai/glm-5.3-flash
Code file edits and commands z-ai/glm-5.3, openai/gpt-5.6-sol
Debug bug diagnostics openai/gpt-5.6-sol
Explore, General codebase exploration, general tasks z-ai/glm-5.3-flash

Saving on sessions

Section titled “Saving on sessions”

Kilo agent sessions send the whole accumulated context with every turn — rules, history, read files. Mixen caches the repeating prefix automatically; nothing to enable.

  • Repeated context is ~10% of the input price: for GPT-5.6 Sol input is 213.9 ₽ per 1M tokens and 21.4 ₽ from cache; for GLM 5.3 Flash — 8.4 and 1.7 ₽.
  • The cache lives 5 minutes with sliding renewal — every request inside the window extends it. A continuous session keeps the cache alive from start to finish.
  • A break longer than 5 minutes — the first turn after it pays the full input. One expensive turn after a pause is the norm, not an error.

Tactics:

  • Planning and reading are cheap turns (input from cache, short output); do them freely and often — save on the model, not on the number of turns.
  • A streak of edits in a row is cheaper than the same edits punctuated by coffee breaks: the cache doesn’t have time to go stale.
  • Streaming is per-symbol — text appears as it’s generated; the price is unaffected.

What a typical session costs

Section titled “What a typical session costs”

Let’s price a reconnaissance session on GLM 5.3 Flash (input 8.4 ₽ per 1M, from cache — 1.7 ₽): 25 turns of ~40K tokens of context each.

Scenario Session input
Without cache 1M tokens × 8.4 ₽ ≈ 8.4 ₽
With cache first turn in full, then the prefix at 1.7 ₽ ≈ 2.6 ₽

A nuance: Flash’s cache discount is shallower than the flagships’ — repeated input costs ~20% of full price (for GPT-5.6 Sol it’s ~10%). But the absolute numbers are so small that long reconnaissance on Flash costs pennies either way; and the cache warmed on it keeps working when you switch to an expensive model in Code — the context is shared. The numbers are indicative: real volumes depend on the project.

The cache share of a response shows up in usage.prompt_tokens_details.cached_tokens — Kilo’s UI doesn’t display it, but when reconciling the bill with the dashboard it explains why a turn cost less than the full input.

Reasoning depth

Section titled “Reasoning depth”

Our API accepts the reasoning_effort parameter: off, low, medium, high, xhigh, or max — the set depends on the model. Allowed values are listed in GET /v1/models → capabilities.reasoning_efforts; the can_disable_reasoning flag shows whether the model accepts off.

  • Reasoning is billed as output tokens and consumes the response-length limit — long tasks need max-token headroom.
  • If the agent or model settings don’t expose a depth knob, the model runs at its default — use the card in the catalog as a reference.

Tips for agentic work

Section titled “Tips for agentic work”
  • Built-in agents. Out of the box: Code (edits), Plan (planning), Ask (questions), Debug (diagnostics), plus Explore and General. The picker remembers the chosen model for each agent separately — set it up once: Plan and Ask on GLM 5.3 Flash, Code on GPT-5.6 Sol.
  • Custom agents. Custom modes are created in Settings → Agent Behaviour → Agents, as markdown files .kilo/agents/*.md with YAML frontmatter, or via the agent key in kilo.jsonc. The mode: subagent property hides an agent from the picker — only other agents can invoke it; primary shows it in the list.
  • Rules. The list of files and globs is set with the instructions key in the project’s kilo.jsonc — for example, ".kilo/rules/*.md"; global rules — in ~/.config/kilo/kilo.jsonc. Format is markdown (recommended) or plain text. Project instructions take precedence over global ones.

A minimal project kilo.jsonc:

{
  // wire up all project rules in one line
  "instructions": [
    ".kilo/rules/*.md"
  ]
}
  • Orchestrator is deprecated. Per Kilo’s docs, Orchestrator mode is deprecated: its scenarios are covered by agents with full tool access (Code, Plan) — delegate to them directly.
  • Tools. All Mixen chat endpoints support tools and tool_choice; for agentic work choose models with tool calling and a large context.

Frugal session checklist

Section titled “Frugal session checklist”
  • Plan, Ask, and Explore — on GLM 5.3 Flash; Code and Debug — on GLM 5.3 or GPT-5.6 Sol.
  • The model is assigned per agent — check that last week’s cheap one didn’t stay on Code.
  • Reconnaissance and planning warm the shared cache — don’t spare the turns.
  • A new task means a new chat: the old history gets paid for on every turn.
  • After refreshing the model list, re-check IDs against the catalog.

Useful documentation

Section titled “Useful documentation”
Section titled “Recommended models”
Model ID
GPT-5.6 Sol openai/gpt-5.6-sol
Claude Opus 5 anthropic/claude-opus-5
GLM 5.3 z-ai/glm-5.3
DeepSeek V4 Pro deepseek/deepseek-v4-pro
Kimi K3 moonshotai/kimi-k3

Full list — in the catalog.

Troubleshooting

Section titled “Troubleshooting”
  • 401 / invalid API key — double-check the mxn-… key in the provider’s API key field.
  • “Model Not Found” — the model wasn’t added to the provider or the ID has a typo; check the catalog.
  • Model list doesn’t load / connection errors — make sure the Base URL is correct (https://api.mixen.ai/v1) and the API is reachable from your network; models can also be entered manually.
  • Rules don’t apply — check the paths in instructions (the key lives in kilo.jsonc, not the UI) and restart the chat session: changes are picked up by a new session.
  • An agent uses the wrong model — the picker remembers the choice per agent; switch the model while inside that agent.