Skip to content

AI assistant

The assistant sees your real project data through tools. It does not guess: before it changes anything it states what it is about to do and asks for your approval.

Providers

Provider Note
Claude, OpenAI, Qwen, DeepSeek, Gemini, Groq Require an API key
Ollama, LM Studio Run locally, no key needed
OpenCode Zen / Go / Free Works with OpenCode credentials
Custom endpoint Any OpenAI-compatible URL

Key, base URL and model are set in the settings (Ayarlar ▸ AI). For OpenCode the credentials can be imported straight from the TUI with Import from opencode TUI. Model lists are fetched live from the provider; the ⟳ button refreshes them.

Your data boundaries

API keys stay on your machine. When you select a provider, the conversation and the tool results are sent to that provider. Use a local provider (Ollama / LM Studio) for sensitive projects.

Permission modes

Mode Behaviour
Ask Every write tool raises an approval card. Default and recommended.
Auto No approval card, but the validation layers still run.
Plan Write tools are locked; the model may only propose a plan.

Tools (61)

They fall into five groups:

Discovery — map search, listing by name, map properties, axes, folders, statistics.

Reading — current and original values, raw ROM reads, similarity search, anomaly detection, shape analysis.

Writing — fill a whole map, write a cell range, scale, offset, clamp, smooth, apply an expression, copy a map, undo.

Management — version snapshots, linked-ROM operations, comments, dyno results.

Planning — plan_start, plan_propose, plan_list, plan_remove, plan_review, plan_finalize, plan_apply.

Plan-first workflow

Put the assistant in Plan mode for large changes. The model follows:

  1. plan_start — opens a session, creates mrc_plan.json and plan.md.
  2. plan_propose — adds a justified proposal per change.
  3. plan_review — lists the proposals so you can drop what you do not want.
  4. plan_finalize — freezes the plan; write tools stay locked.
  5. plan_apply — runs only with confirm=true and only after your approval.

The application takes a snapshot before applying and writes no byte until these checks pass:

  • scaling and byte-order validation,
  • physical value bounds, rejection of NaN/infinite values,
  • rejection of scale factors outside 0.5–3.0,
  • map dimensions versus target ROM size.

Example requests

Find the catalyst monitoring maps in this ROM and zero their thresholds.
Increase the fuel pump map flow by 8%, plan first.
Smooth the high-RPM region of the injection maps.

The first writes (with approval), the second works in plan mode, the third is limited to a single map. The narrower you write the request, the more reviewable the result.

Map name translation

Miscellaneous ▸ Translate Map Names translates map names and descriptions into your target language. The engine chain is Google → Google (Chrome) → MyMemory, and results are cached in the mapnames database.

The live preview table shows Original | Translated columns and the translation column is editable. Remove results you do not like with Miscellaneous ▸ Clear Translation Cache.

Chat history

Chats are stored per project and restored when you reopen it. The history panel lets you return to earlier sessions.

Limits

  • The assistant only reaches the open project's data; it does not roam the file system.
  • Write tools cannot bypass the validation layers.
  • It cannot guarantee that a safety or emissions test will pass — do not decide without measuring.