JAVATEL Inc.株式会社ジャバテル
Home › Resource › SlimeTree-RLM › Meta › Instagram DM Safe Bot

Public MIT

Instagram DM Safe Bot × SlimeTree-RLM Pre-processing

Drop spam DMs before paying the AI tax

For Instagram Business inbound DMs the RLM classifies up-front — spam DMs / investment pitches / identity probes are µ suppressed, FAQs return as D for instant reply, real inquiries go to R for Claude. With Access Token + IG User ID + IGSID you can really Send; without them the simulator runs standalone.

1. Instagram Graph API credentials

At developers.facebook.com/docs/instagram-api/getting-started, link your Instagram Business account to a Facebook Page → add the Instagram Graph API product in the Meta App → issue a Page Access Token.
Required permissions: instagram_business_basic + instagram_business_manage_messages + pages_messaging.

Not authenticated (save token + IG User ID, then click "Check account")
Instagram policy notes:
  • 24-hour window: Free-form replies are allowed only within 24 hours of the customer's last DM. Beyond that, only the HUMAN_AGENT tag or specific message types may be sent
  • Opt-in required: The customer must have actively DM'd your account first (no unsolicited sending)
  • IGSID (Instagram-Scoped User ID): The customer's IG ID is obtained from sender.id in the webhook payload and is different from the normal username
  • IG User ID is not the username: Not @brand but a numeric ID like 17841.... Retrieve it via GET /me
  • Business / Creator accounts only: Personal accounts cannot use the DM API; switching to Business is required

2. Reply scope

Beyond DMs, Instagram also exposes Story replies and Comment replies. This demo focuses on DMs; the others are shown for scope only (Send is implemented for DM only).

DM: POST /v22.0/{ig_user_id}/messages — 1-on-1 free-form reply (within 24h window)

3. Inbound message simulator

Type as the "customer" with a DM → the bot pre-processes via RLM → an automatic reply is generated. Typical examples:
· 商品の値段教えて (Tell me the product price) → D (instant FAQ)
· あ / $$$ 投資案件 $$$ (investment offer) / 住所教えて (tell me your address) → µ (suppressed)
· 御社の事例をフォロワー向けに紹介したい (We'd like to share your case study with our followers) → R (delegated to Claude)

Total received: 0
D: 0
µ: 0
R: 0
Send API: 0
cheap (B): 0
escalated (B): 0
esc rate: —
Send a message below as the "customer" and the bot will respond via the RLM.

3. LLM API key (called only on R verdict)

R-verdict replies are delegated to an LLM (browser-direct). Keys stay in localStorage only and never traverse JAVATEL servers.

GEMINI
CLAUDE · token billing
OPENAI · token billing
LLM-neutral: free choice of Gemini / Claude / OpenAI / Grok — Gemini = free tier available; Claude / OpenAI = each provider's token billing. Fees go directly to your LLM provider; JAVATEL does not mediate.

4.5 Operating mode NEW: Pattern B (cross-vendor) — via the shared module

5. API call log

(nothing has run yet)

Pricing example: 10,000 inbound DMs/month

Monthly cost of the same workload across 6 LLM tiers. The RLM filters 73%, so the reduction ratio is the same on every provider. See the service page for details.

Tier Class No filter With RLM Savings/month Annual
Gemini 2.5 Flash Everyone $7.50 $2.03 $5.47 $66
Gemini 2.5 Pro Everyone $103 $28 $75 $900
GPT-5 mini Business $26 $7 $19 $228
GPT-5 Business $310 $84 $226 $2,712
Claude Haiku 4.5 Business $42 $11 $31 $372
Claude Opus 4.7 Business $1,530 $413 $1,117 $13,404

* 200 in / 1,000 out tokens. "No filter" = all 10,000 DMs go to the LLM; "With RLM" = only the 2,700 R-verdict DMs do. Prices reflect public unit rates as of 2026-05; industry mix is adjustable.

6. Why the RLM works for Instagram DMs (unique strengths)

  • Automatic spam-DM rejection: Influential accounts on Instagram receive a flood of "investment offers", "collab requests", and "affiliate" spam. The µ verdict drops them before the operator reads → prevents staff burnout and fraud exposure
  • D verdict for instant reply inside 24h: Standard inquiries about products / hours / shipping status are 0 tokens → maximizes response rate before the window expires and sustains engagement
  • R verdict for brand-aware replies: Marginal cases (collab pitches, press inquiries) go via Claude for careful responses → reduces brand damage risk
  • Preventing PII leaks: "Tell me your home address" / "Tell me your phone number" are mechanically suppressed by µ → safeguards the influencer's personal safety
  • WAL audit chain: For harassment / report cases, later you can air-gap reproduce "what DM arrived when, and how the bot handled it" via the SHA-256 chain
  • Story replies / Comment replies can be rolled out with the same pattern (this demo implements DM only)

7. Source / Related