qiling-soulmate

Screenshots

Percentages in the settings are estimates or historical ranges for different scopes. They cannot be added together and are not a savings guarantee. See the recorded 5-, 37-, and 246-round cases below for cost comparisons.

Latest maintenance update

• 2026-08-15: Fixed per-turn savings visibility and sub-agent accounting after task completion. Pricing formulas and display precision are unchanged.

A note from the author

I am an independent developer. The AI era gave me a lot of inspiration and made development unprecedentedly efficient; but the high subscription fees and API costs kept bothering me. Every time a new model appeared, I almost had to subscribe to another service, and the old subscription was soon left idle. I also deployed many models locally, but apart from being free, they struggled to be useful in real development work.

So I kept breaking down the differences between strong and weak models on tool calls, file reading and other metrics. In the end I reached a conclusion: let the strong model be the brain — decomposing tasks, judging direction and doing final verification; let the weak model be the hands — file reading, retrieval, repetitive execution and bulk output. I built this Agent CLI, and after countless rounds of testing and rule fixes, this collaboration finally worked.

With the compression rules I designed — I call it 器灵压缩 (Qiling Compression) — the combination of main and collaborative models can greatly reduce the main model's consumption while keeping precision; in some tasks, speed and result quality even exceed the strong model working alone. I also connected the CLIs I commonly use into the same workbench, keeping the collaborative gains while preserving my original habits. If you are a heavy user of AI Agents, you are welcome to join me and raise a Qiling (器灵) of your own.

What it is

器灵 Soulmate (Soulmate) is a local-first AI programming assistant product. Instead of running one assistant at a time, it gathers several Agent CLIs into a single workbench — you pick, route, and observe them side by side, so one channel going down never takes the rest with it.

Its core idea is simple: strong models do the thinking, weak models do the legwork. Strong models handle task decomposition, direction and final verification; weak models handle file reading, retrieval, repetitive execution and bulk output. A dedicated compression layer keeps the strong model's context small without losing precision.

Soulmate runs as a desktop application and works directly on your machine, including browser automation via CDP / Selenium / Playwright.

Choose one route: Fast · Accurate · Economical

Choose exactly one route. Pick the single route that fits your current primary goal — the three are not enabled at the same time, nor run one after another.

Pick one route per task, according to your current primary goal. The numbers below are product-side estimates from the official website scenarios, not customer testimonials or universal promises.

Fast — scout the big repo first, then let the strong model act

In an unfamiliar repository, the task is decomposed into scout ranges, several sub-agents search in parallel, and the strong model receives the condensed key evidence before it starts editing.

• Example (official website scenario): 4 parallel tasks cut a job from 18 minutes to 7 minutes, about 2.6× faster.

Accurate — find counterexamples in parallel, let the expensive model only arbitrate

Before a high-requirement delivery, evidence is cross-checked along separate paths, counterexamples are collected, and the final call is a consensus ruling backed by machine verification (PASS/FAIL) by the strong model.

• Example (official website scenario): scoring goes from 86 to 94 under independent review plus machine verification.

Economical — outsource mechanical steps, keep the expensive model from replaying

Mechanical steps such as retrieval, scanning, drafting and testing are handed to cheaper collaborative models, so the strong model reads less context and executes fewer tool calls.

• Example (official website scenario): cost drops from 100 to 41 cost units, saving 59%.

For conversation baselines, the official website estimates savings of 8–25% for short dialogues (1–3 effective rounds), 35–64% for multi-round dialogues (4–14 effective rounds), and 51–82% for long dialogues (15+ effective rounds). All figures are product-side estimates under fixed assumptions, not guarantees; real results vary with model, pricing, cache and task structure.

Whichever route you pick, remember: Choose exactly one route — align the whole task to your current primary goal, rather than enabling all three or running all of them in turn.

Real Savings Cases

The screenshots below are from real Xiaoyi (internal Qiling build) conversation interfaces. They show per-round cost comparisons on the Economical route, measured against direct use of the original CLI.

Case 1 · Five-turn brief conversation

The entire conversation was only 5 rounds; in round 5, Qiling was outsourced 5 times (4 exploration tasks and 1 coding task) / 77 tool steps → this round dropped from $58.068 to $8.504, saving 85%; cumulative conversation cost dropped from $79.673 to $12.845; cache hit rate was 98%.

Case 2 · Short conversation

A 37-round everyday-task conversation, round 32: Qiling outsourcing 1 time / 11 tool steps → this round dropped from $64.351 to $9.210, saving 86%.

Case 3 · Long conversation

A 246-round large-project conversation, round 186: Qiling outsourcing 6 times (including 3 coding tasks) / 58 tool steps → this round dropped from $271.624 to $21.957, saving 92%; cumulative conversation cost dropped from $2,567.681 to $378.260; cache hit rate was 98%.

![Long conversation real savings case](asse