FormLM

me.formlmv0.5.2更新於 Oct 10, 2026

Build AI-powered forms, scored assessments and professional reports on FormLM, by natural language.

已驗證STDIO僅桌面Data & AnalyticsProductivity & Workflow

概覽

AI 產生的概覽

FormLM 讓助理用自然語言在 formlm.me 上建立 AI 表單、計分測評與專業報告。

功能
此伺服器以 stdio 方式提供 FormLM CLI，內含 9 個工具與 6 個資源。助理可以規劃並產生表單或測評、新增欄位、定義計分維度與分數區間、設定頁面樣式、配置報告頁面與元件、設定用於解讀的 AI 專家，並發佈結果取得分享連結。唯讀的 snapshot 與 doctor 指令可回報模組狀態與品質問題。
適用情境
當你希望助理依描述建立、修改、稽核或發佈 FormLM 表單、計分測評、考試、問卷或報告，而不是透過網頁編輯器操作時使用。也適合在一次工作階段中批次產生並驗證多個應用程式。
執行需求
以 npm 套件（@formlm/cli）透過 npx 在使用者本機執行，需要 Node.js 18 或更新版本。需要 FormLM 帳號，驗證使用存取權杖或聊天內電子郵件驗證碼。可選擇提供 FORMLM_TOKEN 密鑰，FORMLM_BASE_URL 預設指向託管服務。需要能連線至 FormLM 伺服器的網路。
安裝前請注意
此伺服器可建立、修改、發佈與刪除應用程式及其內容，刪除應用程式無法復原。發佈後表單可供匿名訪客填寫，且提交內容永久保留。FORMLM_TOKEN 密鑰可存取帳號，權杖 7 天後過期。啟用 Data API 會暴露提交端點，設定樣式或發佈可能觸發伺服器端 AI 產生。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 FormLM，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

formlm-cli

The official CLI & MCP Server for FormLM — let AI Agents build and manage your forms directly.

[npm version] [License: MIT] [Node.js] [MCP Registry]

Keywords: FormLM CLI, MCP Server, AI Agent form builder, form automation, assessment platform CLI, Claude MCP, Cursor MCP, AI-powered forms, formlm-cli, npm CLI tool

What is FormLM?

FormLM is an AI-powered form & assessment platform. It lets you build smart forms, scoring quizzes, and professional evaluation reports — all with natural language instructions.

With formlm-cli, you can control FormLM directly from your terminal or plug it into any AI Agent (Claude, Cursor, GPT, etc.) as an MCP Server — no UI needed.

Visit the official website: https://formlm.me

formlm-cli is listed in the official MCP Registry as me.formlm/cli — search "formlm" in any MCP-capable client (Claude Desktop, Cursor, VS Code Copilot, Codex CLI) to install it.

What's New in v0.5.0

Hardening round driven by real agent field reports (batch runs of 50 apps), plus a second pass that re-checked every reported item against the code and machine-verified all documented commands.

  • share publish now defaults to anonymous access (--access visitor) and exposes --access / --perm / --days / --no-style; a new share set subcommand passes the raw server parameters through, so "anyone + permanent" no longer requires calling the internal API. Previously publish was hard-wired to form-type all, which requires login — anonymous respondents got the login page.

  • expert config --enable no longer self-destructs: the value is forwarded space-separated (--enable true) and the server accepts it again. Before, --enable=true was rejected as an unknown option, the whole command was discarded, yet the envelope still reported success.

  • Parameter errors are no longer reported as success: picocli validation failures now map to code != 0 / ok:false instead of hiding the error text inside data with code:0.

  • Uniform machine output: snapshot and auth status now honour --json/FORMLM_JSON=1 with the same {ok,code,message,data} envelope as everything else, and human guidance lines moved to stderr, so stdout is a single parseable line.

  • app remove is a real alias of app delete; app list gained --all / --limit / --page / --with-urls; app urls returns shareToken, published, shareType/Perm/Day and absolute https URLs.

  • New doctor command (and MCP formlm_doctor tool): one read-only call per app covering scoring coverage, dead/empty experts, styling, unexpected certificate pages, language consistency (--expect-lang, --deep scans widget text) and share access + reachability.

  • New snapshot --summary compact audit profile (counts, dimension names, report pages, expert flags, share type/perm/day + shareToken) — batch verification no longer needs to parse full module payloads.

  • Resume after the plan cache expires: smart plan --save-plan <file> + smart execute --plan-file <file> (the server plan cache is ~10 minutes, and the docs now say so).

  • Retry/backoff: read-only commands retry with exponential backoff on 408/429/5xx; writes are never auto-retried. Tunable via FORMLM_RETRIES / FORMLM_TIMEOUT_MS.

  • report page remove --name "结业证书" resolves page names server-side; report page update verifies the page exists (opt into upsert with --upsert); write confirmations are compact; widget remove now persists (it previously returned success without saving).

  • smart generate now exists as a resumability-preserving wrapper over plan → execute×N → (--publish) → (--doctor), and the README finally matches the command surface. Earlier versions documented a smart generate that was never implemented (the first command every agent ran, and it failed).

  • Generation is now pin-able: smart plan --dimensions "A|B|C" fixes the scoring dimension names and count, --app-name fixes the app display name, --dry-run validates a prompt without leaving a residue app. Unpinned dimensions used to be renamed/recounted by the AI, forcing page↔app rework (batch measurement: 11/13 pages).

  • Docs are now machine-verified against the command surface: every formlm-cli … example in README/INSTALL is executed and checked for unknown/missing options (149 examples, 0 drifts), so documented commands cannot silently diverge again.

  • smart plan output carries structured modules / missingModules, so callers can detect coverage gaps (e.g. no expert for non-consultation plans) without parsing prose; the raw exec 403 now lists the whitelisted command set.

What's New in v0.2.1

  • Field ID validation: field add --id now enforces snake_case regex (^[a-zA-Z0-9_]+$)
  • Snapshot --md flag: Unified JSON output by default, with optional --md for markdown format
  • Error handler: Missing required options now print usage hints instead of bare error messages
  • Server-side fixes: SessionContext appId sync, ScaleCommand B5 reversal removal, CLI --app extraction

What's New in v0.2.0

The MCP architecture has been completely redesigned from the ground up:

  • 34 flat tools → 6 layered tools + 6 knowledge resources
  • Intelligence Layer: formlm_generate wraps the server-side AssessAgent pipeline
  • Domain Knowledge: 6 MCP resources expose SKILL.md files directly to AI agents
  • State Awareness: formlm_snapshot aggregates all module states in one call
  • P0 Constraints: Embedded directly in command descriptions — AI sees them every time

Architecture

┌─────────────────────────────────────────────────────────┐│         AI Agent (any MCP Client: Claude / Cursor /       ││           Codex CLI / Windsurf / Cline / ...)             │├─────────────────────────────────────────────────────────┤│  Tier 0: auth_login, auth_email_code, auth_status      ││  Tier 1: formlm_generate                   ← Smart Pipeline ││  Tier 2: formlm_snapshot, formlm_skill ← State + Knowledge││  Tier 3: formlm_exec                    ← Direct Commands │├─────────────────────────────────────────────────────────┤│  Resources: formlm://skills/{form,scale,connect,        ││              report,expert,share}                        │├─────────────────────────────────────────────────────────┤│               formlm-cli MCP Server                      │├─────────────────────────────────────────────────────────┤│            FormLM Server (McpV1Api)                      ││  ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐    ││  │AssessAgent│ │BuilderAgent│ │CliService│ │SKILL.md  │    ││  │ (Plan+   │ │ (Think+  │ │ (picocli │ │ (domain  │    ││  │  Execute) │ │  Reflect)│ │  dispatch)│ │  rules)  │    ││  └──────────┘ └──────────┘ └──────────┘ └──────────┘    │└─────────────────────────────────────────────────────────┘

Installation

bash
npm install -g @formlm/cli

Requires Node.js ≥ 18.

For a detailed step-by-step guide (including MCP setup for Claude Desktop / Cursor / Codex CLI / Windsurf / Cline), see INSTALL.md.

Quick Start

1. Login

bash
# Recommended: copy your Access Token from the web app# (formlm.me → sign in → Workspace → top-right user menu → Account Settings → Access Token → Copy)formlm-cli auth login --token <your-token>
# Interactive login — pick one of three methods:#   1) Access Token    (same as above)#   2) Email verification code  (no password, no browser — fully in the terminal)#   3) Email + password (only if your account has set a password —#      email-verification-code and Google accounts have none)formlm-cli auth login

2. Smart Pipeline (AI-recommended)

Creation is a two-phase pipeline (plan → execute per module). smart generate is the one-shot wrapper of the same steps when you do not need per-module control:

bash
# One-shot: plan → execute every module → publish anonymously → quality auditformlm-cli smart generate \  --input "Create a workplace stress assessment with 10 questions, 3 dimensions (workload, autonomy, support), and detailed score interpretations" \  --plan-type assessment --question-count 10-15 --lang en --publish --doctor

Or step by step (recommended for batches — every step is resumable):

bash
# Phase 1 — plan (~10-30s): creates the app and returns appId + task list.# --save-plan keeps the full plan JSON on disk so modules can still be executed# after the server-side plan cache (~10 min) expires.formlm-cli smart plan --input "Create a workplace stress assessment with 10 questions, 3 dimensions, and detailed score interpretations" \  --plan-type assessment --question-count 10-15 --lang en --dimensions "Workload|Autonomy|Support" --save-plan plan.json
# Phase 2 — execute every module listed by the plan, in order (~30-120s each):formlm-cli smart execute --app <appId> --module formformlm-cli smart execute --app <appId> --module scaleformlm-cli smart execute --app <appId> --module connectformlm-cli smart execute --app <appId> --module reportformlm-cli smart execute --app <appId> --module share
# Publish so anonymous visitors can fill it in (each respondent submits once, permanent):formlm-cli share publish --app <appId>            # --access visitor is the default
# Verify quality in one read-only call (see Doctor below):formlm-cli doctor --app <appId> --expect-lang en

⚠️ Module coverage depends on planType. Only consultation plans include an expert task; assessment / exam / report / survey / learn do not. smart plan prints a warning when a module is missing — add an expert explicitly with formlm-cli expert config --app <appId> --name ... --role ... --kbText ....

⚠️ Styling: the connect module of the pipeline applies the visual style. If you skip the pipeline and use Direct Commands instead, you MUST run connect style apply-all, otherwise the form keeps the plain unstyled look.

3. Get All App URLs (after creating)

bash
# Get fill-in, editor, data management, and Data API URLs in one callformlm-cli app urls --app <appId># Returns: shareUrl (relative /s/<token>), shareUrlAbsolute (https, iframe-ready),#          shareToken, published, shareType/sharePerm/shareDay,#          builderUrl (editor), dataUrl (data management),#          apiUrl + apiHelpUrl (Data API, empty until enabled)

4. Enable the Data API (form backend for any page)

bash
# Turn a form into a data endpoint — static sites, local pages, and AI apps# can POST submissions straight to it (CORS enabled, no server needed)formlm-cli share api --app <appId> --submit true --query true# Then point your HTML form / fetch / curl at the returned endpoint.# Append ?help to the endpoint for its Markdown docs (readable by AI agents).

5. Beautify Your Form (if using Direct Commands)

bash
# Apply AI-generated visual style to all pages (takes 30-120s)formlm-cli connect style apply-all --app <appId> --look "职场压力评估，深蓝专业风格" --theme minimalist

6. Direct Commands (for fine-grained control)

bash
# Get a snapshot of all module statesformlm-cli snapshot --app <appId>formlm-cli snapshot --app <appId> --md              # Markdown format (token-efficient for AI)formlm-cli snapshot --app <appId> --summary          # compact audit profile (batch verification)formlm-cli snapshot --app <appId> --with-values      # include report widget body text
# Read a skill document before constructing commandsformlm-cli skill formformlm-cli skill scale
# Scale commandsformlm-cli scale query --app <appId>formlm-cli scale add --app <appId> --id stress --name "Stress Level" --format sum --kbText "..."formlm-cli scale keys add --app <appId> --scale stress --fields q1,q2,q3formlm-cli scale data add --app <appId> --scale stress --bands "Low:desc||Moderate:desc||High:desc"   # recommended: server auto-computes even boundaries + 999 sentinel
# Report commandsformlm-cli report query --app <appId>formlm-cli report page add --app <appId> --name "Summary"                     # page type defaults to A4formlm-cli report widget add --app <appId> --page <pageId> --type scale-chart --format bar --scaleId <dim> --x 0 --y 0 --w 24 --h 12formlm-cli report widget logic add --app <appId> --page <pageId> --id <widgetId> --min 0 --max 20 --content "<p>Keep it up</p>"
# Expert commands (--name and --kbText are required by the server)formlm-cli expert query --app <appId>formlm-cli expert config --app <appId> --name "Career Coach" --role "资深职业规划师" --kbText "..." --theme Warmformlm-cli expert set --app <appId> --property enable --value true   # single-property toggle
# Quality audit (read-only)formlm-cli doctor --app <appId>formlm-cli doctor --app <appId> --expect-lang en --deep

7. Use as MCP Server

bash
formlm-cli mcp

This starts the MCP Server (stdio transport) with 9 tools + 6 resources, ready for AI Agents to connect.

Command Reference

Smart Pipeline (AI-recommended)

There are two ways in: smart plan + smart execute (explicit, resumable — recommended for batches), or smart generate, which is a thin wrapper over exactly those steps:

bash
# Phase 1 — plan only (creates the app, returns appId + tasks; cached server-side ~10min)formlm-cli smart plan --input "..." [--plan-type assessment] [--style "温暖亲切"] [--question-count 10-15] \  [--lang en] [--dimensions "Focus|Sleep|Load"] [--app-name "Team Focus Check"] \  [--save-plan plan.json] [--dry-run]
# Phase 2 — execute one module at a time (form / scale / connect / report / expert / share)formlm-cli smart execute --app <appId> --module formformlm-cli smart execute --app <appId> --module scale --plan-file plan.json   # after the cache expiredformlm-cli smart execute --app <appId> --module scale --clear-first          # re-run without duplicating dimensions
# One-shot wrapper — plan → execute every module → optional publish → optional doctorformlm-cli smart generate --input "..." --plan-type assessment --lang en --publish --doctorformlm-cli smart generate --input "..." --only form,scale,report       # partial buildformlm-cli smart generate --input "..." --skip connect                 # leave styling out

Pinning vs. improvising. By default the Plan AI invents dimension names/count and the app name, which breaks page↔app consistency in batch runs. Use --dimensions to pin the dimension list (exact names + count) and --app-name to pin the display name; verify afterwards with snapshot --summary (it reports dimNames).

--dry-run validates a prompt/spec and deletes the probe app again (the server always creates an app when planning), so experimentation leaves no residue.

Success is machine-readable: smart execute returns data.taskStatus (success / error), and smart generate returns {appId, modules:[{module,status,error}], publish, doctor, resumeHint} — trust those fields, not the human wording. On a mid-run failure the app is kept and resumeHint gives the exact resume command.

Snapshot

bash
# Get all module states in one call (JSON by default)formlm-cli snapshot --app <appId>formlm-cli snapshot --app <appId> --module scale    # Only scale moduleformlm-cli snapshot --app <appId> --md             # Markdown format (token-efficient for AI)formlm-cli snapshot --app <appId> --summary        # Compact audit profile (counts/dim names/report pages/expert/share + shareToken)formlm-cli snapshot --app <appId> --with-values    # + report widget body text (per-page queries)formlm-cli snapshot --apps <id1,id2,...> --summary # Batch: N apps in ONE process (connection reuse, FORMLM_CONCURRENCY parallel apps, default 4)

Measured: 5 apps × 7 module queries = 35 requests in ~13s in a single process, versus one cold Node process per app before (the old serial-subprocess pattern took roughly 5× that, plus no connection reuse).

snapshot is client-synthesized: it fans out to the 6 assess <module> query commands. There is no server-side assess snapshot, so calling that over raw POST /api/v1/mcp/exec returns 403 by design — loop the 6 queries yourself when you need HTTP-side batching. If a module fetch fails, the result carries _errors / _degraded (treat it as unknown, not as "empty module").

Doctor (read-only quality audit)

bash
# One call per app: content/scoring coverage, expert sanity, styling,# unexpected certificate pages, share access semantics, link reachabilityformlm-cli doctor --app <appId>
# Whole batch in one process (aggregate envelope; exit 1 if any app fails)formlm-cli doctor --apps <id1,id2,...> --expect-lang en
# Add the language-consistency scan (classic batch defect: English app shipped with Chinese boilerplate)formlm-cli doctor --app <appId> --expect-lang en            # titles/labelsformlm-cli doctor --app <appId> --expect-lang zh-hant --deep  # also scan report widget body textformlm-cli doctor --app <appId> --no-probe                  # skip the fill-in URL reachability probe

Output is the standard envelope; ok:false + exit 1 when a fail-level finding exists, so batch drivers can gate on it. Each finding carries a ready-to-run fix command. Read-only — it never writes.

Skill Documents

bash
# Get SKILL.md domain knowledge (P0/P1/P2 constraints, command templates)formlm-cli skill form       # Form field rulesformlm-cli skill scale      # Scale dimension rulesformlm-cli skill connect    # Page styling rulesformlm-cli skill report     # Report layout rulesformlm-cli skill expert     # Expert interpretation rulesformlm-cli skill share      # Share/publish rules

Auth

bash
formlm-cli auth login --token <tok>   # Login with Access Token (recommended: formlm.me → Account Settings)formlm-cli auth login                 # Interactive login — 1) Access Token, 2) email verification code (no password needed), 3) email + passwordformlm-cli auth status                # Check current login state (tokens expire after 7 days)formlm-cli auth logout                # Clear local token

Profile (multi-account)

bash
formlm-cli profile add --name work --url https://formlm.me --token <tok>formlm-cli profile listformlm-cli profile use work           # Switch default profileformlm-cli --profile work app list    # Use a profile for one command

App

bash
formlm-cli app list                                               # first page (latest 20)formlm-cli app list --all                                          # full inventory (server page size capped at 500)formlm-cli app list --limit 100 --page 2                           # explicit pagingformlm-cli app list --all --with-urls                              # + published/shareType/sharePerm/shareDay/shareUrl/shareToken per appformlm-cli app list --name "wellness"                              # filter by app nameformlm-cli app create --name "My Form" --description "..."formlm-cli app get --app <appId>formlm-cli app update --app <appId> --name "New Name" --theme blueformlm-cli app delete --app <appId>                                # Delete an app (irreversible)formlm-cli app remove --app <appId>                                # alias of app deleteformlm-cli app urls --app <appId>                                  # Get fill-in / editor / data URLs + shareToken

Field

bash
formlm-cli field schema                                         # All supported field typesformlm-cli field config --type radio                            # Configurable props for a typeformlm-cli field list --app <appId>formlm-cli field find --app <appId> --id q1                     # Get a single fieldformlm-cli field add --app <appId> --id q1 --name "Name" --type input --requiredformlm-cli field update --app <appId> --id q1 --title "Your Full Name"formlm-cli field remove --app <appId> --id q1formlm-cli field set-property --app <appId> --id q1 --property options.0.score --value 5

Scale (Dimensions & Scoring)

bash
formlm-cli scale query --app <appId>                          # List all dimensionsformlm-cli scale find --app <appId> --id <scaleId>            # Find a dimensionformlm-cli scale add --app <appId> --id stress --name "Stress" --format sum --kbText "..."formlm-cli scale update --app <appId> --id stress --name "Stress Level"formlm-cli scale set --app <appId> --id stress --property direction --value negativeformlm-cli scale remove --app <appId> --id stressformlm-cli scale clear --app <appId>formlm-cli scale config --app <appId> --enable-single true
# Associate fields with a dimensionformlm-cli scale keys add --app <appId> --scale stress --fields q1,q2,q3 [--polarities "q3:negative"]formlm-cli scale keys remove --app <appId> --scale stress --fields q1formlm-cli scale keys list --app <appId> --scale stressformlm-cli scale keys set --app <appId> --scale stress --fields q3 --negScore true
# Configure score rangesformlm-cli scale data add --app <appId> --scale stress --bands "Low:desc||Moderate:desc||High:desc"   # recommended: server auto-computes even boundaries + 999 sentinel (ordered lowest→highest score)formlm-cli scale data add --app <appId> --scale stress --ranges "0-10:Low,11-20:High"                   # manual boundaries; only for non-even knowledge-specified thresholdsformlm-cli scale data update --app <appId> --scale stress --id <dataId> --value "Moderate"formlm-cli scale data remove --app <appId> --scale stress --id <dataId>formlm-cli scale data list --app <appId> --scale stressformlm-cli scale data clear --app <appId> --scale stress

Connect (Page Styling & Visual Design)

bash
formlm-cli connect query --app <appId> [--type cover|main|final] [--filter <keyword>] --mdformlm-cli connect find --app <appId> --filter <keyword>formlm-cli connect types [--category cover|main|final] --verboseformlm-cli connect config --category cover --format rich-text
# Cover page managementformlm-cli connect cover-page add --app <appId> --id cover_main --name "Welcome" --format rich-text --value "<h1>Hello</h1>"formlm-cli connect cover-page update --app <appId> --id cover_main --name "New Title"formlm-cli connect cover-page find --app <appId> [--id cover_main]formlm-cli connect cover-page remove --app <appId> [--id cover_main]
# Final page managementformlm-cli connect final-page add --app <appId> --id final_report --name "Thank You" --enable-report trueformlm-cli connect final-page update --app <appId> --id final_report --name "Done"formlm-cli connect final-page find --app <appId> [--id final_report]formlm-cli connect final-page remove --app <appId> [--id final_report]
# Main page managementformlm-cli connect main-page set --app <appId> --format cardformlm-cli connect main-page set --app <appId> --field X1 --description "<p>Context</p>"
# Visual styleformlm-cli connect style set --app <appId> --type cover --bg "#f3f3fe" --fg "#ffffff" --font "#01105c"formlm-cli connect style query --app <appId> --type coverformlm-cli connect style apply --app <appId> --look "deep blue tech, frosted glass cards" [--theme Minimal] [--layout flat]formlm-cli connect style apply-all --app <appId> --look "light warm Japanese style" [--theme Minimal] [--layout flat]formlm-cli connect style move --app <appId> --id <pageId> --direction up

Report (Pages & Widgets)

bash
formlm-cli report query --app <appId>                           # List all report pagesformlm-cli report find --app <appId> --filter <keyword>         # Find report widget
# Note: update goes through the explicit owners — `report page update` (page name/bg/style/svg)# and `report widget update` (widget props). The server also accepts the merged# `assess report update --page <pageId> [--id <widgetId>]` form on the raw exec channel.
# Page management (a page is a 48-col × 68-row grid canvas; layout comes from widget x/y/w/h, not a page flag)formlm-cli report page add --app <appId> --name "Summary" [--type A4]formlm-cli report page update --app <appId> --id <pageId> --name "Overview"      # fails if the id does not existformlm-cli report page update --app <appId> --id <pageId> --name "Overview" --upsert   # opt into the server upsert (creates a blank page)formlm-cli report page remove --app <appId> --id <pageId>formlm-cli report page remove --app <appId> --name "结业证书"                       # locate by human-readable name
# Widget management (every widget command needs --page; set/find/update/remove also need --id)formlm-cli report widget list --app <appId> --page <pageId>formlm-cli report widget find --app <appId> --page <pageId> --id <widgetId>          # full detail incl. HTML valueformlm-cli report widget set --app <appId> --page <pageId> --id <widgetId> --property value --value "{{TotalScore}}"formlm-cli report widget add --app <appId> --page <pageId> --type scale-chart --format bar --scaleId <dim> --x 0 --y 0 --w 24 --h 12formlm-cli report widget update --app <appId> --page <pageId> --id <widgetId> --name "Updated Chart"formlm-cli report widget remove --app <appId> --page <pageId> --id <widgetId>formlm-cli report widget types --verbose                                             # all widget types (+ descriptions)formlm-cli report widget config --type scale-chart                                   # configurable properties of one type
# Conditional display rules (logic lives UNDER widget)formlm-cli report widget logic add --app <appId> --page <pageId> --id <widgetId> --min 0 --max 20 --content "<p>Low band text</p>"formlm-cli report widget logic list --app <appId> --page <pageId> --id <widgetId>formlm-cli report widget logic remove --app <appId> --page <pageId> --id <widgetId> --logicId <logicId>

Expert (AI Interpretation)

bash
formlm-cli expert query --app <appId>                    # Query expert config (--md for a compact table)formlm-cli expert find --app <appId>                     # Full config incl. prompt/welcome textformlm-cli expert config --app <appId> --name "Dr. AI" --role "Career coach" \  --kbText "..." [--welcome ...] [--question1 ...] [--theme Warm] [--enable true]   # create/update (name + kbText required)formlm-cli expert set --app <appId> --property enable --value true    # single-property toggle;                                                                        # properties: name/role/description/style/welcome/                                                                        # prompt/question1..3/kbText/enable/enableWelcome/themeformlm-cli expert avatar --app <appId> --url <url> --enable true      # needs an existing expert (config first)formlm-cli expert remove --app <appId>formlm-cli expert chat --app <appId> --input "Explain my score"

Boolean flags are forwarded space-separated (--enable true), never --enable=true — the server parses them with picocli arity 0..1. expert config enables the agent by default; pass --enable false to configure it while leaving it switched off.

Share (Publish & Access)

bash
# Access types (know these — "public" is ambiguous):#   visitor = anonymous, no login required  → "anyone can fill" (the usual public form)#   all     = every LOGGED-IN FormLM user   → anonymous visitors hit the login page#   secret  = password-protected   |   owner = creator only   |   no = unpublishformlm-cli share publish --app <appId>                      # anonymous (visitor) + one submission + permanentformlm-cli share publish --app <appId> --access all --perm 2 --days 14    # explicit controlformlm-cli share publish --app <appId> --no-style            # skip the automatic style fallbackformlm-cli share set --app <appId> --type visitor --perm 1 --day 0        # raw server parameters (idempotent)formlm-cli share unpublish --app <appId>formlm-cli share query --app <appId>                        # authoritative type/perm/day tripleformlm-cli share verify --app <appId>                       # published + anonymous + permanent + reachability in one callformlm-cli share url --app <appId>                          # URLs incl. shareTokenformlm-cli share api --app <appId>                          # Configure the Data API (form backend endpoint)formlm-cli share api --app <appId> --submit true --query true --auto-create true

Validity rule (server-side): only --days 1..30 are honoured as a finite window; 0, forever, or anything above 30 normalizes to permanent (sentinel 3650000). There is no 90-day/1-year publish window on this channel.

Verifying "anyone can access": curl returning 200 on the share URL proves nothing — the SPA shell answers 200 even behind the login gate. Use share verify (or read share query's type/perm/day triple).

smart execute --module share vs share publish: the pipeline's share module applies the plan's publish settings; share publish is the explicit, idempotent CLI entry that also guarantees anonymous + permanent defaults and prints the final URLs/shareToken. Re-running share publish is safe. If you never styled the app, publish auto-applies a default style first (30-120s AI generation) — pass --no-style to skip that.

The Data API turns a form into an HTTP data endpoint: POST JSON to the endpoint (or use the /form path for native HTML forms), read records back via the query endpoint, and hand the ?help URL to an AI agent so it can discover the API on its own.

MCP Integration

FormLM CLI works as a standard MCP Server over stdio and plugs into any MCP-compatible AI platform — Claude Desktop, Cursor, Codex CLI, Windsurf, Cline, etc.

⚠️ macOS/Linux users: desktop AI clients often launch the MCP server without your full terminal PATH, which can cause command not found errors. See INSTALL.md → Step 4 for the env.PATH fix, per-platform config file locations (including Codex CLI's TOML format), and the full setup guide.

No token? No problem. If you omit FORMLM_TOKEN, the AI will prompt you to authenticate via the auth_login tool — paste your Access Token (from formlm.me → Workspace → Account Settings; recommended), complete an in-chat email verification-code login via auth_email_code (no browser needed), or use email + password directly in the chat.

Available MCP Tools (9)

TierToolDescription
0auth_loginLogin with Access Token (recommended), email verification code, or email + password
0auth_email_codeEmail verification-code login — fetch the captcha image & send the code (fully in-chat)
0auth_statusCheck current login status
1formlm_generateGenerate the execution plan (Phase 1: creates the app, returns appId + tasks)
1formlm_executeExecute plan modules one by one after formlm_generate
2formlm_doctorRead-only quality audit of one app (scoring coverage, dead experts, styling, certificate pages, language consistency, share access, reachability)
2formlm_snapshotAggregated state of all modules; summary: true returns the compact audit profile
2formlm_skillFetch SKILL.md domain knowledge for a skill module
3formlm_execExecute any whitelisted CLI command directly

Available MCP Resources (6)

Resource URIDescription
formlm://skills/formForm field SKILL.md — P0 constraints, field types, scoring rules
formlm://skills/scaleScale dimension SKILL.md — range rules, polarity, direction
formlm://skills/connectConnect page SKILL.md — theme modes, layout formats
formlm://skills/reportReport layout SKILL.md — widget types, variables, logic
formlm://skills/expertExpert interpretation SKILL.md — AI config, chat rules
formlm://skills/shareShare/publish SKILL.md — access types, idempotency

Environment Variables

VariableDescription
FORMLM_BASE_URLFormLM server URL (default: https://formlm.me)
FORMLM_TOKENYour auth token (alternative to auth login; keeps secrets out of shell history)
FORMLM_JSON=1Machine output: every command prints ONE parseable line {ok,code,message,data} on stdout (equivalent to the global --json flag)
FORMLM_NO_EXIT=1Batch safety: never process.exit() on failure — a bad command cannot kill your driver loop (errors surface as ok:false envelopes / thrown errors)
FORMLM_TIMEOUT_MSDefault per-command request timeout (ms, default 60000)
FORMLM_TIMEOUT_PLAN / FORMLM_TIMEOUT_EXECUTE / FORMLM_TIMEOUT_STYLEOverride the smart-plan (120s), smart-execute (300s) and style (600s) timeouts
FORMLM_RETRIESMax attempts for transient failures (default 3). Read-only commands retry on 408/429/5xx; writes are never auto-retried
FORMLM_CONCURRENCYIn-process parallelism for batch commands (snapshot --apps, doctor --apps), default 4

Scripting & Batch Use

bash
# One parseable line per command; guidance/hints go to stderr, stdout stays clean:FORMLM_JSON=1 FORMLM_NO_EXIT=1 formlm-cli snapshot --app <appId> --summaryFORMLM_JSON=1 formlm-cli share verify --app <appId>        # {ok, data:{type,forever,anonymous,httpStatus,shareToken}}FORMLM_JSON=1 formlm-cli doctor  --app <appId> --expect-lang en   # ok:false when a fail-level finding exists
# Whole batch in one process (no per-app Node cold start; tunable with FORMLM_CONCURRENCY):FORMLM_JSON=1 formlm-cli doctor   --apps id1,id2,id3 --expect-lang enFORMLM_JSON=1 formlm-cli snapshot --apps id1,id2,id3 --summary

Envelope: { "ok": bool, "code": int, "message": string, "data": object|null } — data is always parsed (no double-decoding needed), and validation/parameter errors now come back with a non-zero code instead of hiding the error text inside data while reporting success.

Status codes

CodeMeaningWhat to do
0Success—
207Partial (multi-module reads where one module failed)Result carries _errors/_degraded; re-check that module, don't assume it is empty
400Bad request / parameter or validation errorRead message; the server echoes picocli's hint (valid values, missing option)
401Not authenticatedformlm-cli auth login --token-stdin (or set FORMLM_TOKEN)
403Command not on the exec whitelist, or wrong credentials/verification codeSee the whitelist below; for MCP check auth_status
408Request timed out client-sideRetry read-only; for smart execute re-run that module (writes are not auto-retried)
429Rate limited (login attempts, send-code quotas)Wait, then retry with backoff
500Transport failure or server errorRetried automatically for read-only commands

Raw POST /api/v1/mcp/exec whitelist

The exec channel accepts 3-token command paths only (assess <module> <action>). Everything else returns 403 not allowed via MCP by design — notably assess snapshot and assess field … do not exist on the server (they are client-side wrappers), so batch HTTP consumers must call the underlying queries:

assess app      list | create | use | current | update | remove | urlsassess form     query | find | types | config | add | update | remove | move | set-propertyassess scale    query | find | add | update | set | remove | clear | config | keys | dataassess connect  query | find | types | config | cover-page | final-page | main-page | styleassess report   query | find | update | page | widget        (conditional logic is `report widget logic …`)assess expert   query | find | config | set | avatar | remove | chatassess share    set | query | url | api | flavorassess smart    plan | executeassess skill    form | scale | connect | report | expert | share

Throughput notes

  • Every CLI invocation is a fresh Node process plus its own HTTP calls; for large batches, run independent apps concurrently (the server tolerates moderate parallelism) rather than serially.
  • snapshot/doctor fan out their module queries in parallel; use --module / --summary to keep payloads small.
  • Slow server? Raise FORMLM_TIMEOUT_MS and FORMLM_RETRIES instead of hand-rolling retry loops.
  • Batch over many apps inside one process: snapshot --apps … / doctor --apps …, or import the built module directly (import { execCommand } from '@formlm/cli/dist/exec.js') and loop — HTTP connections are reused and there is no per-app Node cold start.

Version coupling (CLI ↔ server)

Some commands rely on server options added together with them; against an older server they fail with a visible 400 Unknown option … (never a silent success, thanks to the false-success guard):

Needs server supportCommand / flag
✔app list --all / --limit / --page / --with-urls
✔app urls native shareToken + published/shareType/… (the CLI fills shareToken from shareUrl as a fallback)
✔report page remove --name "…" (name→id resolution moved server-side, so raw exec benefits too)
✔expert config --enable true|false (fall back to expert set --property enable on an old server)
✔smart plan --dimensions / --app-name, smart execute --plan-file (--plan-b64 transport)

Links

Contact

Have questions, feedback, or need help getting started?

📧 [email protected]

Feel free to reach out — we're happy to help.

License

MIT © FormLM

來源：README.md，提交 0e7b839

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v0.5.2最新Oct 10, 2026