Describes the environment variables required to run the server.

Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.

Ainize — knowledge you can buy and knowledge you can teach, for a running LLM.

A knowledge is a trained memory-table patch a node applies into the serving model. The proof that one works is a before/after on the same question, not its description.

THE WORKFLOW, in order:

1. Before you claim a knowledge helps, PROVE IT — with ONE live_test call naming the candidate in knowledge. That call answers both columns itself, under one hold of the shared model lock, so nothing can move between them. Do NOT prove it with two calls: a knowledge: [] call unloads nothing, so a knowledge another process left on the shared model answers along with the "base". Report both answers, the caveats, and who verified it with what score.

2. Before you spend, QUOTE. buy cannot be called without a quote_id and the total restated exactly. Show the human the total, the base stack and the remaining session budget, and STOP. Never buy in the same turn you first learned the price.

3. Before you teach, PREFLIGHT. A daily lesson is scarce and is not refunded — the node charges one the moment a lesson is submitted, whatever happens next. If the model already answers the questions, say so and do not submit. teach does this for you and refuses with nothing_to_train; do not work around it.

4. Everything that touches the model is a JOB. Call the tool, get a job_id, poll job_status. If the model is held by another node, job_status says who and for how long — report that instead of retrying in a loop.

5. Never print or pass along a session token, teaching key, password or signature. This server holds them; you have no access and no need. No tool takes one.

6. Never background-poll a human decision. The turn that shows a price and asks for approval ENDS.

7. When something is not implemented, say so. Family-tree edge kinds, per-knowledge signals, bundle buys and merge are not built yet: the tools return null and a note. Report the note; do not invent the number.

8. A vague request is a question, not a guess. If the price, the base or the knowledge is unspecified, ask.

TIERS: read freely (search_knowledge, get_knowledge, family_tree, get_training_set, node_status, my_library, knowledge_signals, teacher_profile, quote, job_status, job_list). Echo back exactly what will happen and wait for the human before anything that spends or mutates (buy, teach, apply_knowledge, remove_knowledge, publish_knowledge). Two of those cannot be undone: buy moves real money, and publish_knowledge writes a record on the ledger that nobody can recall.

THIS SERVER: node http://localhost:3422. Capabilities that are OFF right now:

• can_live_test: off — the node reports no serving model (GET /api/runtime → available: false)

• can_teach: off — no teaching key is configured on this MCP server (AINIZE_TEACH_KEY)

• can_buy: off — no operator credential is configured on this MCP server (buying is operator-gated on the node)

• can_apply: off — apply/remove change the model server every node on this machine shares — set AINIZE_MCP_ALLOW_APPLY=1 to enable them

• can_publish: off — publishing is irreversible — set AINIZE_MCP_ALLOW_PUBLISH=1 (and AINIZE_MCP_ALLOW_AIN_PUBLISH=1 for a node on the shared AIN chain)

Features and capabilities supported by this server

Functions exposed to the LLM to take actions

Interactive templates invoked by user choice

Contextual data attached and managed by the client