Skip to main content
Glama

stemseal (lite)

Small LLMs guess at calculations. stemseal gives them verified cells to call instead. 작은 LLM은 계산을 어림짐작합니다. stemseal은 그 대신 부를 수 있는 검증된 세포를 줍니다.

A cell is one small, deterministic function. It enters the bank only after it passes a purity check (no I/O, no clock, no randomness) and at least 20 ground-truth cases. Then it is locked by sha256. A locked cell never changes. A fix becomes a new version. If no cell fits, the model gets an honest "not available" instead of an invented answer.

세포는 작고 결정적인 함수 하나입니다. 순수성 검사와 정답 사례 20개 이상을 통과해야 들어오고, sha256으로 잠깁니다. 잠긴 세포는 바뀌지 않으며, 고치면 새 버전이 됩니다. 맞는 세포가 없으면 지어낸 답 대신 "없음"을 정직하게 돌려줍니다.

This lite version runs entirely on your machine. There is no server, no account, no dependencies, and no telemetry. 이 판은 전부 내 컴퓨터에서 돕니다. 서버, 계정, 의존성, 원격 수집이 없습니다.

What we measured / 측정 결과

52 held-out tasks, 3 runs each (156), local models through Ollama:

Model

Alone

With stemseal + model guide

EXAONE 3.5 7.8B

23%

78%

qwen2.5:3b

11%

66%

llama3.2:3b

7%

50%

Caveats, honestly / 솔직한 한계:

  • One test set, written by us, mostly Korean business arithmetic. Your tasks will differ.

  • 우리가 만든 시험 하나, 주로 한국어 업무 계산입니다. 여러분의 일과는 다를 수 있습니다.

  • Cells help only where a cell exists. Writing, translation and judgment stay with the model.

  • 세포가 있는 계산에서만 돕습니다. 글쓰기, 번역, 판단은 모델 몫입니다.

Related MCP server: math-logic-mcp

Install / 설치

pip install stemseal
stemseal verify          # re-checks every lock and ground-truth case

Connect to your model (MCP) / 모델에 연결

Any MCP client (LM Studio, Claude Desktop, Ollama front-ends that speak MCP, ...):

{
  "mcpServers": {
    "stemseal": {
      "command": "stemseal-mcp",
      "env": { "STEMSEAL_MODEL": "qwen2.5:3b" }
    }
  }
}

STEMSEAL_MODEL is optional. When set, find_cell returns a short selection guide written for that model family's known weak spots (too many candidates, wrapped arguments, renamed keys), and call_cell repairs those input mistakes before running the cell. STEMSEAL_MODEL을 넣으면 그 모델 계열의 약점에 맞춘 고르기 안내가 함께 오고, 흔한 입력 실수를 고친 뒤 실행합니다.

Tools: find_cell, call_cell, list_cells, request_cell.

Language: guides answer in Korean when the request contains Hangul and in English otherwise. Set STEMSEAL_LANG=ko or en to fix it. The numbers above were measured with the Korean wording; the English wording is a translation and has not been re-measured yet. 안내 문구는 요청에 한글이 있으면 한국어, 없으면 영어로 나갑니다. 위 수치는 한국어 문구로 잰 것이고, 영어 문구는 아직 따로 재지 않았습니다.

Command line

stemseal list
stemseal find 연평균 성장률
stemseal call currency_convert '{"amount": 100, "rate": 1385.5}'
stemseal request 대출 원리금 균등상환 월 납입액
stemseal demand

Cells

cell

computes

inputs

annual_leave_days

근속 연수로 연차 일수 (15일, 2년마다 +1, 최대 25)

years

currency_convert

외화를 원화로 환산 (원 미만 버림)

amount, rate

food_calories

100g당 열량과 먹은 양으로 섭취 열량

kcal_per_100g, grams

geometric_mean_growth

기간별 성장률의 기하평균 성장률 (%)

rates

monthly_wage_from_hourly

시급을 월 209시간 기준 월급으로

hourly_wage

percentage_point_diff

두 비율의 차이 (%p)

before_pct, after_pct

percentile_rank

값이 목록에서 몇 % 위치인지

values, value

price_for_target_margin

원가와 목표 마진율로 판매가

cost, margin_pct

Need a cell that isn't here? / 필요한 세포가 없나요?

When find_cell comes back empty, the model can call request_cell. The need is saved only on your machine (~/.stemseal/ledger.jsonl), and the reply carries a link to open a GitHub issue. Nothing is sent unless you open it. Requested cells are built as demand gathers and ship in the next version.

request_cell은 필요를 내 컴퓨터에만 기록하고, GitHub 이슈를 여는 링크를 돌려줍니다. 링크를 직접 열기 전에는 아무것도 보내지 않습니다. 요청이 모이는 순서대로 세포를 만들어 다음 버전에 담습니다.

Safety

Cells run in a separate python -I process with a timeout, and only after the purity gate. This is process isolation, not an OS sandbox. Every cell here was reviewed before locking, and stemseal verify lets you check that none changed.

License

MIT. See LICENSE.

Available Tools

4 tools
call_cellA

Run a locked cell found with find_cell (use its exact name; do not guess names). Use the cell's required input keys exactly, as in its input example. Pass its input either inline as input or, for large data, as input_file (path to a JSON file holding the input object) so the data need not be retyped. Returns status ok (verified result), not_available, or invalid_input. Never invents a result. If refused, the reply's how_to_fix shows the exact input to send again.

Inputs: cell (required), input, input_file.

find_cellA

Call this first whenever the user asks for a number worked out by rules: wages, salary, rates, growth, margins, prices, percentiles, calories, currency, leave days and similar. Verified cells give the exact checked answer, and small models get these wrong when computing alone. State the need in plain words. Returns the cells that fit, each with its required inputs and an input example, or an empty list when nothing fits.

Inputs: need (required), words.

list_cellsA

List every locked cell on this machine with what it computes and its required inputs. Cheap; use it when unsure what exists.

Inputs: no named parameters.

request_cellA

Tell the hub that you needed a calculation and no cell fit. This is how new cells get made: you get a ticket, and the cell is built as demand gathers, so come back next time. Describe the need in one or two sentences.

Inputs: need (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a distinct role in the workflow: list_cells browses all cells, find_cell searches by need, call_cell executes a known cell, and request_cell asks for missing functionality. The descriptions clearly separate discovery, execution, and demand requests, so misselection is unlikely.

    Naming Consistency5/5

    All names follow a consistent verb_noun snake_case pattern: list_cells, request_cell, find_cell, call_cell. The only variation is natural pluralization on list_cells, which is not confusing.

    Tool Count5/5

    Four tools is well-scoped for a verified-calculation hub: browse, search, execute, and request. Each tool has a clear purpose and earns its place, with no obvious redundancy or thinness.

    Completeness4/5

    The surface covers the core lifecycle: see available cells, find one matching a need, run it, and request a new one when none fits. A direct get-cell-by-name lookup independent of semantic search is a minor gap, but agents can work around it via find_cell or list_cells.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers