msds-chain-mcp

Chemical safety intelligence for AI-assisted experiment design.

Powered by ChainSDS — a verified, always-current chemical safety database. Verified. Current. Growing.

An MCP server that gives AI agents (Claude Code, Cursor, Copilot, pi, etc.) access to chemical safety reasoning — compatibility checks, hazard analysis, regulatory compliance, PPE recommendations, storage guidance, and more.

Built for researchers who design experiments with AI and need safety verification integrated into their workflow.

Why This Exists

When you use Claude to plan a synthesis route or set up an Opentrons protocol, safety validation shouldn't be a separate step. This MCP server lets your AI assistant automatically:

• Check if chemicals on the same deck are compatible
• Flag dangerous mixing orders
• Recommend PPE for the specific chemicals you're handling
• Verify compliance with EU REACH, US OSHA/TSCA, and 7 other jurisdictions
• Generate signed audit reports for GLP/GMP compliance

Tools (23)

Tool	Description
batch_safety_check	One-call report for a chemical list: pairwise compatibility + per-chemical risk warnings
check_regulatory_lists	Cross-reference a chemical against 23 regulatory watch lists across 8 jurisdictions + 3 international conventions
get_sds_section	Retrieve a specific SDS section (1-16) for a chemical
get_sds_document	Signed download URL (~5 min) for the original SDS/MSDS PDF; includes source provenance
get_chemical_alternatives	Safer substitutes for restricted or high-risk chemicals
validate_protocol_chemicals	Extract & validate chemical names from protocol text or code
check_mixing_order	Safe addition sequence for reagent pairs (e.g., acid into water)
get_waste_disposal	Waste classification, container type, and disposal procedures
upload_msds_pdf	Upload MSDS PDF for AI-powered parsing and data extraction (requires API key)
compare_sds_versions	Hazard-change diff between two SDS versions (H-code additions/removals + whether they change a verdict)
check_chemical_compatibility	Pairwise compatibility for 2+ chemicals
get_chemical_risk_warnings	GHS classification, H-codes, signal words, flash point
get_ppe_recommendation	Gloves, eye protection, respiratory, body protection
get_storage_guidance	Storage class, cabinet type, temperature, isolation rules
get_emergency_response	Spill, fire, or exposure emergency procedures
get_exposure_limits	OEL/TLV/PEL/MAC across US, EU, JP, CN, INT
get_transport_classification	UN number, hazard class, packing group, ADR/IATA/IMDG
check_regulatory_compliance	Multi-region: EU, US, CN, JP, CA, AU, SG list-backed; KR/TW accepted but no list wired
search_chemical_database	Look up chemicals by name, synonym, or CAS number
ask_chemical_safety	Natural language catch-all for any safety question
create_audit_session	Full audit with signed PDF report (requires API key)
get_audit_report	Download link for the signed audit PDF

Quick Start

1. Get an API key

Sign up at msdschain.lagentbot.com → API Keys tab → create a key.

2. Install

git clone https://github.com/littleblakew/msds-chain-mcp.git
cd msds-chain-mcp
pip install -r requirements.txt

3. Add to your AI coding agent

Architecture note: the hosted endpoint https://mcp.lagentbot.com sits behind a distribution gateway that terminates OAuth 2.1 (PKCE + DCR) and serves /.well-known/oauth-authorization-server. The OAuth / browser sign-in flow described below applies only to that hosted gateway. This repo's bare core server (python server_remote.py, or the Docker image) does not implement OAuth or serve any /.well-known endpoint — it only accepts Authorization: Bearer sk-msds-... or X-Api-Key passthrough. If you self-host the core directly, use a static API key; there is no browser sign-in step.

Claude Code (Remote — recommended):

claude mcp add msds-chain --transport http https://mcp.lagentbot.com/mcp

Then run /mcp in Claude Code and authenticate — a browser opens to sign in (email code); your account is provisioned automatically and calls are metered to you (free plan includes a monthly free quota). The legacy SSE endpoint (--transport sse https://mcp.lagentbot.com/sse) still works but streamable HTTP is preferred.

Claude Code (Plugin — includes skill + MCP):

/plugin install https://github.com/littleblakew/msds-chain-mcp.git

Claude Code (npm — remote, shim registers the hosted endpoint):

claude mcp add msds-chain -- npx -y msds-chain-mcp@latest

claude.ai (Web): Search "msds-chain" in Settings > Plugins (already published).

Manual config (Claude Code ~/.claude.json):

{
  "mcpServers": {
    "msds-chain": {
      "type": "http",
      "url": "https://mcp.lagentbot.com/mcp"
    }
  }
}

Restart Claude Code, run /mcp, and authenticate (OAuth). You should then see msds-chain in the MCP tools list. (Prefer a static key instead of OAuth? Pass it as a header — --header "Authorization: Bearer sk-msds-your-key" — for headless clients.)

Claude Code Skill

The /msds-safety-check skill provides auto-detection and guided audit workflows.

Plugin install (includes skill + MCP automatically):

/plugin install https://github.com/littleblakew/msds-chain-mcp.git

Manual install (skill only, if MCP already configured):

git clone https://github.com/littleblakew/msds-chain-mcp.git /tmp/msds-chain-mcp
cp -r /tmp/msds-chain-mcp/skills/msds-safety-check .agents/skills/msds-safety-check
ln -s ../../.agents/skills/msds-safety-check .claude/skills/msds-safety-check

What the Skill Does

• Auto-detects chemicals in your conversations and offers safety checks
• /msds-safety-check — guided audit for lab protocols or EHS compliance
• Freemium — basic queries work without API key, audit reports require free registration

Usage Examples

Experiment Protocol Review

User: I'm planning a Grignard reaction with magnesium turnings, diethyl ether,
      and bromobenzene. Check if this setup is safe.

Claude:
  → calls batch_safety_check(["magnesium", "diethyl ether", "bromobenzene"])
  → Returns: compatibility matrix + per-chemical risk warnings (with source SDS)

Opentrons Deck Safety Audit

User: My Opentrons deck has these in different slots:
      Slot 1: Acetone, Slot 3: Concentrated H2SO4, Slot 5: Methanol,
      Slot 7: Sodium borohydride. Any safety issues?

Claude:
  → calls check_chemical_compatibility(["acetone", "sulfuric acid", "methanol", "sodium borohydride"])
  → ⚠️ INCOMPATIBLE: Sodium borohydride + sulfuric acid (violent reaction, H2 gas evolution)
  → ⚠️ CAUTION: Acetone + sulfuric acid (exothermic)
  → Recommendation: Move sodium borohydride to maximum distance from acids

Compliance Check Before Shipping

User: We need to ship toluene and dichloromethane to our Japan lab.
      What transport regulations apply?

Claude:
  → calls get_transport_classification(["toluene", "dichloromethane"])
  → calls check_regulatory_compliance(["toluene", "dichloromethane"], regions=["JP"])
  → Returns: UN numbers, IATA packing instructions, Japan-specific regulations

Generate Signed Audit Report

User: Create a safety audit report for our quarterly review.
      Chemicals: acetone, methanol, ethanol, isopropanol, hexane.

Claude:
  → calls create_audit_session("Q2 2026 Solvent Cabinet Review", ["acetone", "methanol", "ethanol", "isopropanol", "hexane"])
  → calls get_audit_report("SESSION-ID")
  → Returns: Signed PDF URL (Ed25519 signature, suitable for GLP/GMP compliance)

Try These Prompts (one per tool)

Each prompt reliably triggers the named tool. Use them to evaluate the connector end-to-end.

Tool	Example prompt	What a good response shows
search_chemical_database	"Look up the chemical with CAS 67-64-1."	Acetone identified with synonyms + CAS
ask_chemical_safety	"Is it safe to store bleach next to ammonia?"	Plain-language hazard answer (toxic chloramine gas)
check_chemical_compatibility	"Are acetone and hydrogen peroxide compatible?"	Pairwise verdict + reaction risk
get_chemical_risk_warnings	"What are the GHS hazards of toluene?"	H-codes, signal word, flash point
get_ppe_recommendation	"What PPE do I need to handle concentrated sulfuric acid?"	Gloves/eye/respiratory/body guidance
get_storage_guidance	"How should I store sodium hydroxide and nitric acid?"	Storage class, cabinet, isolation rules
get_emergency_response	"What do I do if I spill chloroform?"	Step-by-step spill procedure
get_exposure_limits	"What's the OEL for benzene in the EU?"	OEL/TLV/PEL values by region
get_transport_classification	"How is acetone classified for air freight?"	UN number, hazard class, packing group
check_regulatory_compliance	"Is dichloromethane restricted in Japan and the EU?"	Per-region compliance status
check_regulatory_lists	"Which regulatory watch lists include formaldehyde?"	Matched lists across regions
get_waste_disposal	"How do I dispose of waste acetonitrile?"	Waste class, container, procedure
check_mixing_order	"What's the safe order to mix sulfuric acid and water?"	Correct addition sequence + why
get_chemical_alternatives	"Is there a safer substitute for hexane in extraction?"	Lower-hazard alternatives
validate_protocol_chemicals	"Validate the chemicals in this protocol: [paste text]"	Extracted + verified chemical list
get_sds_section	"Show me section 4 (first aid) of the SDS for methanol."	Requested SDS section content
compare_sds_versions	"What changed between SDS versions for acetone?"	Hazard changes between the two versions
batch_safety_check	"Run a full safety check on acetone, methanol, and hexane."	Compatibility matrix + risk warnings
upload_msds_pdf	"Parse this MSDS PDF and extract its hazard data." (API key)	Structured extraction from the PDF
create_audit_session	"Create a signed audit report for our solvent cabinet." (API key)	Audit session + signed PDF
get_audit_report	"Give me the download link for that audit report." (API key)	Signed PDF URL

Privacy & Data Handling

• Privacy Policy: msdschain.lagentbot.com/privacy
• Read-only tools send only the chemical names / queries you provide; no personal data is required.
• API-key tools (upload_msds_pdf, create_audit_session, get_audit_report) associate activity with your account for audit traceability.
• Uploaded MSDS PDFs are processed for data extraction and contribute to the verified ChainSDS database per your account's data-sharing settings.

Third-Party AI Platform Integration

Connect to the hosted MSDS Chain MCP server from any AI platform that supports MCP.

Server URL (preferred): https://mcp.lagentbot.com/mcp (streamable HTTP)

The hosted endpoint serves both transports — streamable HTTP (/mcp, preferred) and SSE (/sse, for clients that only speak SSE: 悟空, Dify, Coze, etc.). Calls are metered per authenticated user.

Transport	Endpoint	Notes
Streamable HTTP	https://mcp.lagentbot.com/mcp	Preferred — broadest, most robust
SSE	https://mcp.lagentbot.com/sse	For SSE-only clients

Authentication: either OAuth (the client opens a browser sign-in on first connect) or a static header — Authorization: Bearer sk-msds-your-key — for headless clients.

悟空 (Wukong)

1. 设置 → MCP 服务 → + 添加
2. 类型：SSE
3. 名称：msds-chain
4. URL：https://mcp.lagentbot.com/sse
5. HTTP Headers：Authorization : Bearer sk-msds-your-key
6. 点击「添加」

Dify / Coze / other platforms

General steps:

1. Find MCP server / tool integration settings
2. Select SSE transport type
3. Set URL to https://mcp.lagentbot.com/sse
4. Add authentication header: Authorization: Bearer sk-msds-your-key

Get an API key: Sign up at msdschain.lagentbot.com → API Keys tab → Create Key.

---

Remote Mode (HTTP)

For cloud deployment or shared team access, run as an HTTP server:

MSDS_API_KEY=sk-msds-xxx python server_remote.py

A single server_remote.py process always serves both transports concurrently — Streamable HTTP at /mcp (recommended for Claude Code 2026+) and SSE at /sse (compatibility). The MSDS_MCP_TRANSPORT env var is kept for backward compatibility only; it is ignored at runtime.

Connect from Claude Code:

claude mcp add msds-chain --transport http https://your-server.example.com/mcp

Docker

docker build -t msds-chain-mcp .
docker run -p 8080:8080 -e MSDS_API_KEY=sk-msds-xxx msds-chain-mcp

Configuration

Variable	Default	Description
MSDS_API_KEY	(required)	API key from the MSDS Chain dashboard
MSDS_API_URL	Production URL	For development use only
MSDS_LANG	en	Response language: en, zh, ja, de, id

Use Cases

Lab Automation (Opentrons / Hamilton / Tecan)

• Pre-run protocol safety audit
• Deck layout compatibility verification
• Automated run compliance reports

Electronic Lab Notebooks (Benchling / LabArchives)

• Safety annotations on experiment entries
• Chemical registration with auto-flagging

Pharmaceutical R&D

• Synthesis route safety screening
• Multi-region regulatory pre-checks for new compounds
• GMP-ready audit trail with signed receipts

Data Coverage — ChainSDS

Industry-sourced, AI-verified, and cryptographically signed.

• 4,000,000+ chemical records with multi-language aliases (EN/ZH/JA)
• NFPA/GHS classification for compatibility rules
• 23 regulatory watch lists across 8 jurisdictions + 3 international conventions (EU, US, CN, JP, KR, CA, AU, SG; no Taiwan and no IARC coverage)
• Occupational exposure limits from 5 standards (OSHA PEL, ACGIH TLV, EU IOELV, JP OEL, CN GBZ)
• UN transport data for 16+ common lab chemicals
• Version tracking — detects a newer SDS and diffs its hazard classification against the one you relied on

Architecture

Your AI Agent                  MSDS Chain MCP Server
┌──────────────────┐           ┌─────────────────────────┐
│ Claude Code      │           │ 23 Safety Tools         │
│ Cursor / pi      │──MCP────▶│   ↓                     │
│ Any MCP client   │           │ ChainSDS Platform       │
└──────────────────┘           └─────────────────────────┘

The MCP server connects to the ChainSDS platform for verified safety reasoning.

Development

Test locally with the MCP inspector:

export MSDS_API_KEY=sk-msds-your-key
npx @modelcontextprotocol/inspector python server.py

Troubleshooting

Symptom	Likely cause	Fix
401 Unauthorized on connect	Missing or invalid API key	Verify the Authorization: Bearer sk-msds-... header (or MSDS_API_KEY env var). Generate a fresh key at msdschain.lagentbot.com → API Keys.
Tools don't appear in the client	MCP server not loaded	Fully restart the client after adding the server. Confirm msds-chain shows in the MCP/tools list.
Connection drops or times out	Wrong transport / endpoint	Prefer the streamable HTTP endpoint https://mcp.lagentbot.com/mcp (SSE sessions can drop when the server redeploys). Check GET https://mcp.lagentbot.com/health returns {"status":"ok"} — that is the gateway; the core's tool count lives on an internal endpoint that is not publicly reachable, so a missing tools field is expected.
Quota exceeded / 429	Monthly call limit hit	Free keys are limited; upgrade your plan or wait for the monthly reset.
Empty results for a known chemical	Name not matched	Retry with the CAS number or a common synonym; search_chemical_database accepts name, synonym, or CAS.
OAuth login fails in a browser client	OAuth metadata unreachable (hosted gateway only — self-hosted core has no OAuth)	Confirm GET https://mcp.lagentbot.com/.well-known/oauth-authorization-server returns 200; the client should auto-discover the authorize/token endpoints. If self-hosting server_remote.py directly, skip OAuth and use a static Authorization: Bearer key instead.

Still stuck? Email contact@lagentbot.com or open an issue at github.com/littleblakew/msds-chain-mcp/issues.

Roadmap

• get_waste_disposal — waste classification and disposal guidance
• check_mixing_order — safe addition sequence for reagent pairs
• get_chemical_alternatives — safer substitutes for restricted chemicals
• Remote MCP (HTTP SSE / Streamable HTTP) for cloud-hosted access
• OAuth 2.1 for Claude Marketplace integration (served by the distribution gateway in front of mcp.lagentbot.com, not by this core server)

License

MIT

About

Built by LAgentBot — AI-powered chemical safety infrastructure.

Part of the MSDS Chain platform — the world's first AI Agent-driven chemical safety data trust network, powered by ChainSDS: verified, current, and growing.