Live read-only no account
Use the graph from your assistant
Ask ChatGPT, Claude or your editor about a supplement, a compound or an outcome, and it can answer from the graph instead of from memory: each finding with its evidence strength, the sentence it was drawn from, the paper, and a link back to the finding here so you can check it.
The server address
https://catalystproject.ai/mcp
- Transport
- Streamable HTTP
- Key · sign-in
- none
- Tools
- 7, every one read-only
- Install on this side
- nothing
- asked“does melatonin help sleep?”
- called
find_evidencequerymelatonin
· aboutsleep
- returnedeach finding with its evidence strength in words, the study design, the quote, the paper, a permalink and a citation to quote as it stands
- answeredwith the permalink,
/e/{id}, so the reader can check it. If nothing matches, it says Catalyst has no findings on it.
Paraphrase of the tool description below. Not a recorded answer.
§ 1
Connect it
Claude
In Claude (web or desktop): Settings → Connectors → Add custom connector, name it Catalyst and paste the address above. Then turn it on for a conversation from the tools menu.
ChatGPT
Until Catalyst is listed in ChatGPT’s plugin directory, ChatGPT can connect it as a custom connector in developer mode, using the address above. Which plans offer developer mode is ChatGPT’s to decide and changes; its help pages say.
Claude Code
claude mcp add --transport http catalyst https://catalystproject.ai/mcpVS Code and GitHub Copilot
Install in VS Code opens the editor and asks before adding it. Or put this in .vscode/mcp.json:
{"servers": {"catalyst": {"type": "http","url": "https://catalystproject.ai/mcp"}}}
Cursor
Install in Cursor, or add this to ~/.cursor/mcp.json:
{"mcpServers": {"catalyst": {"url": "https://catalystproject.ai/mcp"}}}
Gemini CLI
gemini mcp add --transport http catalyst https://catalystproject.ai/mcpWindsurf, Cline, and other MCP clients
Most clients configured by a JSON file accept this block:
{"mcpServers": {"catalyst": {"url": "https://catalystproject.ai/mcp"}}}
§ 2
What it can look up
7 tools, listed here exactly as the server describes them to the assistant.
- findingsmeasured, with evidence strength
- hypothesesinferred from reference databases
- lookupfinds ids, says nothing about them
- 01
find_evidencefindingsFind the evidence on a substance or a question
Answers "what does X do?" and "does X affect Y?" for a supplement, nutrient, drug, food compound or plant (for example "does magnesium help sleep?", "what is creatine shown to do?", "omega-3 and triglycerides").
Takes
query, and optionallyabout,evidence_strength,limit,detailThe whole description the assistant reads
Answers "what does X do?" and "does X affect Y?" for a supplement, nutrient, drug, food compound or plant (for example "does magnesium help sleep?", "what is creatine shown to do?", "omega-3 and triglycerides"). Pass the substance as query and the effect or outcome asked about, if there is one, as about. When the person names two substances ("is creatine safe with caffeine"), pass one as query and the other as about, in one call; findings about the pair together come back under in_combination. It resolves the name to the node that has findings and returns up to limit findings (default 10), strongest evidence first, each with its evidence strength in words, the study design, the quote, the paper, a permalink and a citation to quote as it stands. total_findings counts every reported finding that matches; hypotheses counts the inferred rows listed beside them. The result is complete for the query: another wording of the same substance resolves to the same node, so do not call again to rephrase. Call again only to narrow (evidence_strength, about), to raise limit (at most 50), or to ask about something else. Without about, summary lists everything the node has findings about. If matched is null, or total_findings is 0, Catalyst has recorded no findings on it. That is a complete answer: say so, and do not answer from memory as though from Catalyst. Try search_nodes only when other_matches or summary shows a node the person plausibly meant. An evidence strength says how well the evidence supports a finding and is not a recommendation. For what a compound can become, use reach; for reference-database relations, use get_relations.
- 02
search_nodeslookupFind a compound, outcome or protein
Looks up the node id for a compound, protein, organism or outcome by name or synonym (for example "CoQ10", "caffeine", "CYP1A2", "sleep").
Takes
queryThe whole description the assistant reads
Looks up the node id for a compound, protein, organism or outcome by name or synonym (for example "CoQ10", "caffeine", "CYP1A2", "sleep"). Use it to get the id that get_findings, get_finding, get_relations, get_reactions and reach take; for a question about what a substance does, call find_evidence instead, which needs no id. Returns up to 10 hits, each with id, label, type and the number of findings; hits with findings come first, so the first hit with findings above 0 is usually the one to use. An enzyme or transporter has type protein. It says nothing about a node beyond its id and how many findings it has. If with_findings is 0 the graph holds no findings for the name; the hits may still be the right node for get_relations, get_reactions or reach.
- 03
get_findingsfindingsFindings about a node, with their evidence strength
Lists the curated study findings for one node by id, from search_nodes: for a compound, organism or material, what it has been shown to affect; for an outcome, which compounds have been studied against it.
Takes
node_id, and optionallylimit,about,evidence_strength,detailThe whole description the assistant reads
Lists the curated study findings for one node by id, from search_nodes: for a compound, organism or material, what it has been shown to affect; for an outcome, which compounds have been studied against it. Use it to browse a node by id. For a question in words, such as "does X help Y", find_evidence answers in one call and needs no id. summary has one row per thing the node has findings about, counted over all its findings and not only the rows returned. The default number of findings, strongest first, is enough to answer. There is no offset: to see a different slice, pass about (an id from summary, or words such as "sleep") or evidence_strength. Each finding carries its evidence strength, one of five named levels from strong to insufficient (strong, moderate, limited, very_limited, insufficient), the verbatim sentence it was drawn from, the paper, the population and dose, and a permalink. An evidence strength says how well the evidence supports the finding and is not a recommendation; give it with the claim, and link the permalink. An empty list means no findings are recorded for that node, a complete answer that does not show the compound does nothing; empty_means says why.
- 04
get_findingfindingsOne finding, with its evidence strength explained
Explains one finding by its finding id: the claim, its evidence strength and what that level means, the verbatim quote, the study design and the paper, and evidence_strength_reasons, each rule that set the level in order, and what the assessment does not weigh.
Takes
finding_idThe whole description the assistant reads
Explains one finding by its finding id: the claim, its evidence strength and what that level means, the verbatim quote, the study design and the paper, and evidence_strength_reasons, each rule that set the level in order, and what the assessment does not weigh. The finding id is 26 characters, from the id field of a find_evidence or get_findings finding, or the last part of its permalink (/e/{id}). Use it to answer why the evidence behind a finding is as strong as it is, from evidence_strength_reasons and never by guessing. To find findings, use find_evidence or get_findings; this tool takes no name. A head-to-head comparison id returns a sentence naming the page that shows the comparison, not a finding.
- 05
get_relationshypothesesReference-database relations (hypotheses)
Lists what public reference databases (Reactome, UniProt, Rhea, ChEBI and others) record about one node by id, from search_nodes: transporters, enzymes, pathways, expression.
Takes
node_id, and optionallypredicate,limit,offsetThe whole description the assistant reads
Lists what public reference databases (Reactome, UniProt, Rhea, ChEBI and others) record about one node by id, from search_nodes: transporters, enzymes, pathways, expression. Every row is assertion_class 'inferred': a hypothesis about how a compound could act, not a result anyone measured in people. Present these as possible mechanisms and never as effects. For effects, use find_evidence. For the reactions a compound or enzyme takes part in, use get_reactions; for what a compound can become, use reach. predicate limits the list to one relation type, named as in the predicate field of a row (for example transported_by for transporters, metabolized_by for enzymes). Returns total and up to limit rows (default 10); offset pages the rest, and the default is enough to answer. A page too large for one result is cut to fit, and truncated with next_offset says how to get the rest. An empty list with a predicate means the databases record no relation of that type for this node. That is a complete answer, so do not repeat the call with the same predicate. Leave predicate out once only when the person asked broadly and you need to see which types the node has.
- 06
get_reactionshypothesesReactions a compound or protein takes part in (inferred)
Lists the biochemical reactions from Rhea and Human-GEM that a compound is a substrate or product of, or that a protein (enzyme) catalyses or transports, by node id from search_nodes.
Takes
node_id, and optionallylimit,offsetThe whole description the assistant reads
Lists the biochemical reactions from Rhea and Human-GEM that a compound is a substrate or product of, or that a protein (enzyme) catalyses or transports, by node id from search_nodes. It lists reactions and does not answer a route question: for "can X become Y" or "what can X become", use reach. Every row is assertion_class 'inferred': a reaction two public databases record, not a result anyone measured in a person. Present these as known biochemistry, never as an effect a compound has, and never as a recommendation. For effects, use find_evidence. Returns total and up to limit reactions (default 10), each with its equation, its participants by side, id and label, the first five enzymes with catalyst_count for all of them, and its source; offset pages the rest, and the default is enough to answer. A page too large for one result is cut to fit, and truncated with next_offset says how to get the rest. A compound flagged is_hub (water, ATP, protons and the like) returns its reaction count only, because it takes part in nearly every reaction. Coverage is one Rhea release and one Human-GEM release; a participant that is not a node in this graph is listed with a null id, and only human enzymes that are nodes are listed. An empty result means no reaction in those releases names this node. That is a complete answer and does not show the body has none.
- 07
reachhypothesesWhat a compound can become, and through what (inferred, a hypothesis)
Answers a route question: "can X become Y in the body" (give from_id and to_id) and "what can X become" (give only from_id).
Takes
from_id, and optionallyto_id,depth,limit,offsetThe whole description the assistant reads
Answers a route question: "can X become Y in the body" (give from_id and to_id) and "what can X become" (give only from_id). Both ids are compound node ids from search_nodes, so a question that names two compounds needs one search_nodes call for each first. It searches the reactions two public databases (Rhea, Human-GEM) record, up to depth steps (1 to 3, default 3), and returns the routes with the reactions and enzymes along each. Every route is assertion_class 'inferred': a hypothesis drawn from reference databases, never a measured or reported effect. Present it as known biochemical connectivity, never as an effect, a recommendation, or a claim that the body actually does this. For measured effects, use find_evidence. Common cofactors and currency species (water, ATP, protons and the like) are not used as intermediate steps, so a route never reads 'reaches everything through ATP'. A currency species given as from_id or to_id returns reachable: null with a coverage note, not a checked 0 routes. With from_id alone it lists up to limit routes (default 10, offset pages the rest); a page too large for one result is cut to fit, and truncated with next_offset says how to get the rest. When no route is found, the coverage field says which of three cases it is: reachability has not been built for this scope yet, so nothing was checked; it was checked and no route exists through the reactions loaded, which does not mean the body cannot make it; or the compound is a currency species and was never searched. The answer is in that one result; another tool does not check it again.
§ 3
What every answer carries
An evidence strength, and what it means
Every finding comes with the strength of its evidence, on five named levels from strong to insufficient, computed from the study behind it and never set by hand. It says how well supported the finding is. It is not a recommendation to take anything, and nothing here is medical advice.
The sentence and the paper
The verbatim sentence the finding was drawn from, the paper, the population, the dose studied where the paper gives one, and a permalink to the finding on this site, so the person reading the answer can check it against its source.
Findings and hypotheses, kept apart
A connection assembled from reference databases is marked as a hypothesis in the data itself, and the assistant is told to call it one. It never arrives looking like a reported result.
Nothing for sale
No tool returns a product, a price, a plan or a link to buy anything, and no tool ranks what to take. The only link an answer carries is back to the evidence.
How evidence strength is computed, with every ceiling and penalty, is on how it works.
§ 4
Limits
Calls are counted per minute, per client and network address. ChatGPT and Claude each have 600 a minute, because one address there carries many people; Gemini has 300; editors have 120; any other client 60. A program that needs more can use a key (see access). The same graph is readable over plain HTTP at the API and as files from the export.
| ChatGPT | 600 |
|---|---|
| Claude | 600 |
| Gemini | 300 |
| VS Code | 120 |
| Cursor | 120 |
| Any other client | 60 |
§ 5
What is recorded
Not what you asked, with one exception: when a question finds nothing, the server adds one to a daily count against its wording, if the wording is a short name such as “zinc” and says nothing about the person asking. The count carries the day and nothing else, and is deleted after a year; it tells us what to add to the graph next. It also counts how many calls each tool served each day, with nothing attached but the day and, for a call made with a paid API key, a number standing for that key. There is no account to attach anything else to. To enforce the limits above, the server counts calls per minute against the client’s name and a keyed hash of the network address, never the address itself. The rest is in the privacy notice.
Questions, or something an assistant got wrong from this graph: hello@catalystproject.ai.