Key features
- Claude and ChatGPT query the H2-bench DB directly
- Source excerpt, page and DOI in every answer
- No installation
Requirements
- A claude.ai account (free works, allows 1 custom connector) or a paid/Edu ChatGPT account
- A Google account
- Your real name
How to connect
The screenshots below show the Korean claude.ai UI. English labels may differ slightly.
-
1
Path: Settings → Customize → Connectors → "+ Add" → "Add custom connector"
① Connectors ② "+ Add" ③ "Add custom connector" · real screenshot, Korean UI · personal info blurred -
2
Fill in the fields → "Continue"
- Name
- h2-bench
- MCP server URL
- https://h2-mcp.hydrochan.com/mcp
https://h2-mcp.hydrochan.com/mcp① Name ② MCP server URL ③ "Continue" · real screenshot, Korean UI -
3
The connector now appears in the list → "Connect" → sign in with Google and approve in the browser (see B4)
-
4
Chat input + → "Connectors" → confirm the h2-bench switch is on
① Connectors menu ② switch · real screenshot, Korean UI, personal info blurred · shown under whatever name you gave it (the example account shows h2-mcp)
- Free plan: limited to 1 custom connector
- Team/Enterprise: the org Owner registers it under Organization settings → Connectors → Add → Custom → Web; members just click Connect
- Claude Chrome extension side panel (Max/Team now, Pro rolling out): the same connector works there unchanged
-
1
Add the server
claude mcp add --transport http h2-bench https://h2-mcp.hydrochan.com/mcpAdd
--scope userto use it in every project. -
2
In Claude Code, run
/mcp→ h2-bench → Authenticate -
3
The same Google sign-in and approval flow opens in your browser (see B4)
Plans: Plus, Pro, Business, Enterprise, Edu — not available on Free · web only.
-
1
Left sidebar "Plugins" → top right "Add" → "Create MCP App"
① Plugins ② Add ③ Create MCP App · real screenshot -
2
Enter Name and Server URL; leave Authentication as OAuth (default) → check "I understand and want to continue" → Create
- Name
- h2-bench
- Server URL
- https://h2-mcp.hydrochan.com/mcp
① Name ② Server URL ③ "I understand and want to continue" · real screenshot -
3
Google sign-in and approval in the browser (no capture after Create; same flow as B4)
-
4
Below the chat input, "Plugins" → select h2-mcp from the list
① Plugins ② h2-mcp · real screenshot · personal info blurred
Screens shown are from a ChatGPT Edu workspace — menu names and visibility can vary by plan and workspace settings. Not available on the Free plan.
Gemini web: custom apps (MCP) are currently limited to US, English-language, personal accounts — the menu is absent on Korean accounts (checked 2026-09-26).
Sign-in and approval
- 1Check the title
- 2Check the email
- 3Enter your real name (1–40 characters)
- 4"I confirm my affiliation and approve" → done
- Any Google account works (work or personal)
- What's recorded: real name, email, approval time, and tool-call history (tool name and arguments)
- Token values are never recorded
- Reconnecting: the access token renews automatically every hour; the refresh token lasts 30 days, so you sign in again roughly every 30 days
Available tools (10)
The AI selects tools automatically; you only type the question. The table lists what each tool covers.
| Tool | Question it answers | Scope |
|---|---|---|
list_corpora | Which corpora exist, and how many papers/rows each | All |
corpus_stats | Scale and distribution (reaction class, active metal, support, year, etc.) | All; some fields NH3-only |
search_benchmark | Individual experimental results (rows) matching conditions, with source excerpts | All |
top_performers | Rankings by metric (unified basis, converted values shown, review-pending rows excluded) | All; LOHC carrier-pair filter |
get_paper | All rows and limitations for one paper (looked up by DOI) | All |
find_paper | Find a paper by title, author, journal, or year | Across all corpora at once |
search_fulltext | Find mechanisms/explanations in the paper body and SI | Corpus must be specified |
get_paper_fulltext | Read one paper's body text in order (up to 12 chunks) | Corpus must be specified |
query_argmap_claims | Find a paper's key claims | NH3 only, for now |
get_argmap_links | Support/attack/qualify relations between claims | NH3 only, for now |
Example questions
How many corpora are in this database, and how many papers in each?
list_corpora
3 corpora
- LOHC: 391 papers · 4,479 rows · loaded 2026-09-24
- NH3: 1,250 papers · 10,776 rows · loaded 2026-08-13
- SBH: 310 papers · 4,216 rows · loaded 2026-08-13
In NH3 decomposition, show top catalysts by hydrogen generation rate at or below 450°C, with reaction conditions, as a table
top_performers · corpus=nh3 · metric=h2_formation_rate · temp_max=450 · basis=per_gcat
- Basis: mmol gcat⁻¹ h⁻¹ (normalized per gram of catalyst)
- 116 of 468 candidate rows converted from metal loading
- 33 rows excluded by the default filter (reaction_class=thermocatalytic)
| Rank | Catalyst | Loading | H2 formation rate [mmol gcat⁻¹ h⁻¹] | Temp | Source |
|---|---|---|---|---|---|
| 1 | Ru/Sm2O3 | 1.0 wt% | 8,292 | 450 °C | Xu et al., Nature Communications (2024) |
| 2 | Ru/Y2O3 | 1.1 wt% | 7,494 | 450 °C | Xu et al., Nature Communications (2024) |
| 3 | Ru/CeO2 | 1.0 wt% | 6,942 | 450 °C | Xu et al., Nature Communications (2024) |
| 4 | Ru/Gd2O3 | 0.8 wt% | 6,204 | 450 °C | Xu et al., Nature Communications (2024) |
| 5 | 1.0Ru/CeO2 | 1.0 wt% | 5,954 (converted) | 450 °C · GHSV 228,000 mL gcat⁻¹ h⁻¹ | Hu et al., Applied Catalysis B (2020) |
| Ru/Sm2O3 | 1.03 | 300,000 | 138.2 | 13417 | This work |
Original value 138.2 mmol H2·gcat⁻¹·min⁻¹ → server-normalized to 8,292 mmol gcat⁻¹ h⁻¹.
Show experimental results reporting NH3 conversion for Ru/Al2O3 catalysts between 400–500°C, sorted by citation count
search_benchmark · corpus=nh3 · active_metal=Ru · support=Al2O3 · reaction_class=thermocatalytic · temp_min=400 · temp_max=500 · sort=citations
3 of 96 results shown
- David et al., JACS (2014): Ru-Al2O3 · 500 °C · conversion 82% · NH3 60 sccm · evidence p.3 body · verified. [summary] compares NaNH2 90%, Ru 82%, and Ni 58% at the same temperature.
- Wang et al., ACS Sustainable Chem. Eng. (2019): Ru/Al2O3 2.12 wt% · 400 °C · conversion 1.57% · H2 31.5 mmol/(gcat·h) · GHSV 30,000 mL/gcat·h · evidence p.2 body · verified. However, the metal-loading evidence (table, p.5) failed its check → check the original.
- Di Carlo et al., Int. J. Hydrogen Energy (2014): commercial Ru/Al2O3 8.5 wt% · 500 °C · conversion 99.7% · Ea 117 kJ/mol · evidence p.4 body · verified.
Find what the paper text says about why alkali promoters like K raise Ru catalyst activity
search_fulltext · corpus=nh3 · query=potassium promoter effect on Ru activity · top_k=3
Response metadata: backend=hybrid_origin · reranked=false (rrf)
- Li et al., Sustainable Chemistry and Pharmacy (2024), p.7–8: [summary] the promoter donates electrons to Ru, weakening N–Ru adsorption and easing the rate-determining step — N recombination and desorption.
- Hanifah et al., Applied Catalysis B (2026): [summary] on Ru/MgAl2O4, the K promoter raises activity but trades it off against stability.
In the debate over structure sensitivity (B5 sites) from Ru particle shape and size, who supports or attacks whom?
query_argmap_claims · corpus=nh3 · front=structure-sensitivity-B5 get_argmap_links · corpus=nh3 · doi=10.1039/d4sc06382a
94 claims · 11 links for this paper (the first 4 are relation=support, confidence 0.82–0.86)
- [summary] Karim: B5 site count depends strongly on particle shape and grows with size up to about 7 nm.
- [summary] Chen: extends this to carbon-supported Ru that turns into stepped nano-pyramids during the reaction.
Show top catalysts by conversion for DBT/H18-DBT dehydrogenation
top_performers · corpus=lohc · metric=conversion · carrier_pair=DBT/H18-DBT
Default reaction_class=dehydrogenation excludes 44 rows — 21 remain on the % basis
| Rank | Catalyst | Loading | Conversion | Temp | Source |
|---|---|---|---|---|---|
| 1 | Pt-Sn/Al2O3-450 | 0.7 wt% | 98.3% | 290 °C | Peng et al., ACS Applied Nano Materials (2023) |
| 2 | Pt-Sn/Al2O3-550 | 0.7 wt% | 97.4% | 290 °C | Peng et al. (2023) |
| 3 | Pt-Sn/Al2O3-600 | 0.7 wt% | 95.8% | 290 °C | Peng et al. (2023) |
| 4 | 1 wt% Pt/Al2O3 | 1.0 wt% | 90% | 355 °C | Modisha & Bessarabov, Sustainable Energy & Fuels (2020) |
| 5 | 1 wt% Pt/Al2O3 [stability] | 1.0 wt% | 85% | 340 °C | Modisha & Bessarabov (2020) |
| Pt-Sn/Al2O3-450 | 98.3 | 8100 | 71.5 |
Show the data extracted from paper 10.1038/s41467-024-49981-9
get_paper · corpus=nh3 · doi=10.1038/s41467-024-49981-9
Xu et al., Nature Communications (2024) · 43 rows · class thermocatalytic + computational
- "Ru/Sm2O3" row: metal loading 1.03 wt% → p.48 table (verified)
- space velocity 300,000 NH3 mL gcat⁻¹ h⁻¹ → p.48 table (verified)
- temperature 450 °C → p.48 table (verified)
- pure NH3 feed → p.7 body (verified)
- catalyst 50 mg → p.7 body (verified)
- another row: TOF 3.2 s⁻¹ (400 °C, p.4 body) · apparent activation energy 102.2 kJ/mol (p.4 body)
Tips and notes
Tips
- Name the corpus: "in NH3 decomposition" · "in LOHC" · "in SBH"
- Give a temperature range, e.g. "450°C or below" — avoids skew toward high-temperature results
- Say "experimental results only" to exclude computational (CFD/DFT) rows
- For noble-metal comparisons, ask for both per-catalyst-gram and per-metal-gram tables
- For LOHC, specify the carrier pair (e.g. DBT/H18-DBT) and dehydrogenation vs. hydrogenation
- Catalyst names work in standard notation (Ru/Al2O3)
Notes
- A number without a source excerpt is suspect — check the original
- A "(converted)" tag marks a value the server derived (e.g. from loading) — it differs from the number printed in the paper
- A "check failed" tag means you should check the original directly
- Full-text search results are body excerpts, not structured rows — recheck numbers via row lookup
- LOHC rows marked review-pending are excluded from rankings
- Call limits: 20 per minute and 200 per day per account (UTC day · resets daily at 09:00 KST)
- On
rate_limited_minute, wait a minute and retry · onrate_limited_daily, retry after 09:00 KST