T-knowledge is a store for an assistant's knowledge in which every line is a principle, a checked claim (T) or an assumption (A), signed by a named person, with the ids it rests on. It is reached as an MCP server.
https://mcp.tknowledge.org/mcpno-reply@tknowledge.org). Paste it in. Your first sign-in opens a store of your own.claude mcp add --transport http tk https://mcp.tknowledge.org/mcp
Then /mcp, pick tk, and sign in.
Connect to https://mcp.tknowledge.org/mcp with the header Authorization: Bearer <key>. For a key, write to quocbao.tn@icloud.com; it comes with a private link that holds the key on its own — treat it like a password. Not yet tried on Cursor itself.
| Kind | Meaning |
|---|---|
| claim | A statement about the world that can be checked against a source. Every T and every A is one. |
| P.n | A principle: an axiom, added and retired, never rewritten. A new store starts with seven. |
| T-n | A claim that passed the gate — a fact backed by a document, or a conclusion from at least two standing entries. Its wording never changes. |
| A-n | An assumption, with what would settle it. Promoted to T when the evidence arrives, refuted when the opposite is observed. |
| D-n | A document: readable text plus a pointer to the original (which stays where you keep it), and how you looked at it. Bound when a URL and hash tie the text to an original, unbound when it is text alone — both are valid evidence. |
| TH-n | A thread: one matter you follow, with keywords, linking the entries, people and conversations that belong to it. |
| I-n | An instruction: how an assistant should use this store — a meta line (M1), a directive (1, 2 …) or a definition (def:claim). Not a claim and not a principle; added or retired, never edited. |
| S-n | A summary of a thread: always an A, with a named author; each sentence cites ids in the thread; marked stale when newer entries arrive. |
| conv_n | A conversation: a chat or message trail linked to a matter, kept as context — it informs, it is not cited as proof. |
| F-n | A flag: something needs a person — evidence withdrawn, a source that changed between readings. The store raises it instead of resolving it quietly. |
| attester | A person who stands behind a claim by adding their own document for it. |
The assistant picks these on its own; you talk in plain language.
| Tool | Use |
|---|---|
| assume · update_assumption | Hold something not yet checked, and revise it while it is open. |
| document_add | Record evidence as text, with where the original is. |
| admit · promote | Put a checked claim into the ledger; raise an assumption once its evidence is in. |
| attest · withdraw_evidence | Add or withdraw a document behind a claim. |
| demote · retire · recheck · refute | Take down what no longer holds; record that a source was checked again; refute an assumption. |
| principles · principle_add · principle_retire | Read, add or retire principles; retiring one lists everything standing on it. |
| instructions · instruction_add · instruction_retire | How assistants use your store, kept in the store: one list for every project — meta lines about the list, numbered directives, and definitions of the store's concepts. Added or retired, never edited; wording already there is caught as for claims. |
| store_info | Who maintains this deployment and where to send feedback when tk misbehaves. |
| thread_open · thread_tag · link · unlink · thread_close | Group entries into a matter you follow. |
| summarize | Write a reading of a thread — drafted by the assistant, approved by you, every sentence citing ids. |
| ledger · entry · chain · open_items · thread · lookup · documents · flags | Read: what is known, one entry, what it stands on, what is still open, a thread, search by keyword, documents, warnings. |
| check_claim · define | Ask whether a sentence is a claim yet, and what the store's words mean. These work before signing in. |
A refusal is an answer, not an error. It comes with a code, a sentence saying why, and next — the move still open — so the assistant can tell you what to write instead. Common ones:
CLAIM_FAILED — not a claim yet: two sentences in one, a relative date, a person not named.NO_SOURCE — a fact names no document, or not where in it the claim is read.SINGLE_PREMISE — a derivation needs at least two premises.PREMISE_NOT_FOUND — a premise id is not in the truth ledger (an assumption is not).USE_PROMOTE — that sentence is already an open assumption; promote it with evidence.SOURCE_NOT_RETRIEVABLE, USE_DOCUMENT — a fact needs a document someone else could open.PREMISE_NOT_T — a conclusion cannot stand on something that has fallen.SUMMARY_UNCITED — a sentence in a summary cites no entry of the thread.CLAIM_TOO_MANY_WORDS, WRITE_LIMIT_REACHED — a store's limits.Every refusal carries its own reason and next. Ask the assistant to run define on a word of the store — claim, T, A, evidence, promote … — for the store's own explanation.
no end date, 1000 writes, reads unlimited; up to 100 words per claim. When the writes run out the store stays readable.