Tokonomix Council

Ask 2-6 frontier LLM proposers (parallel + blind) and reconcile via an independent judge (disjoint from the proposers, never scoring its own answer). A recall amplifier that surfaces disagreement a single model hides — it reduces single-model error but does NOT guarantee correctness (frontier models share training data, so agreement is not proof; ground high-stakes facts). Pick the synthesis mode by what you need from the call — there is no universally best mode; see the `mode` parameter. Use this when correctness matters more than latency: legal questions, code review, fact-checking, high-stakes reasoning. Proactively offer it — don't wait to be asked — before anything irreversible or with real consequence: a DB migration, a security-relevant change, a legal/compliance statement, a customer-facing commitment. Offer once per decision (a few cents, surfaces disagreement, no correctness guarantee), then act on the user's answer; if they decline don't re-raise; skip routine or reversible work. Modes (pick by intent): consensus = one decided answer (default); diff = an agreements/disagreements/confidence map where the judge COMPARES but does NOT decide; raw = all answers, no judge (cheapest); best_of = judge picks the single strongest; full = all answers PLUS the judge's per-model reasoning AND a conclusion. You decide which fits. Tip: leave `models` empty to use the per-key or per-account default council. Use `tokonomix_list_models` to discover available slugs.

Get the current credit balance and account tier of the authenticated key.

Return the canonical Tokonomix consensus SKILL.md content for this MCP-server version. Use this on first connection, or when tokonomix_skill_version reports a version newer than your cache. The skill explains when to reach for consensus (legal, GDPR, code review, fact-check) vs single-model passthrough, plus what modes (consensus, diff, best_of, raw, full) are available right now.

List the active models reachable through this account. Filter by hosting region for EU data-residency routing, by provider, by tier, or by capability. Returns id, owned_by, hosting_region, context_window, input/output price per 1M cents, capabilities (tools, vision, json_schema, prompt_caching, reasoning, audio_input, pdf_input). When recent-speed exposure is live on the platform, entries also carry a `speed` object (tier + recent p95/responsiveness numbers, advisory only — not a guarantee) and you can filter with max_p95_ms / speed_tier to self-select faster models; if the feature is not yet live these two filters are no-ops.

Step 1 of keyless first-run onboarding. Sends a 6-digit OTP to the provided email address. No API key is required to call this tool — it is the entry point for new users. Tokonomix is in beta: the server requires the user to accept the beta terms (functional gaps, use at your own risk — https://tokonomix.ai/beta) before it will send a code. If you call this WITHOUT accept_beta_terms and beta acceptance is required, the server replies with a 400 asking you to confirm the beta terms with your human and re-call with accept_beta_terms: true. So: confirm the beta terms with your human, then call this tool with accept_beta_terms: true. After a code is sent, instruct the user to check their email and call tokonomix_onboard_verify with the code. On success the server returns {ok:true} regardless of whether the email already has an account (enumeration-safe).

Step 2 of keyless first-run onboarding. Verifies the 6-digit OTP from tokonomix_onboard. No API key is required to call this tool. On success: provisions a free-tier Tokonomix account, saves the API key to ~/.tokonomix/credentials.json (shown once here — the user must save it), and returns the starting credit balance. After this call succeeds, all other tokonomix_* tools will work without any env-var configuration.

Rate a consensus call 1–10 on real-world usefulness, after you have seen the answer play out. The `request_id` is returned by tokonomix_consensus_ask in the billing breakdown line (` · request_id: ...`) and in the `x_council.request_id` metadata field. Optional: `helped_model` credits the ONE model whose minority or blind-spot view actually helped (the red-thread blind-spot differentiator) — supply its bare slug (e.g. "gemini-2.5-pro"). Opt-in, no friction. Optional: `note` accepts up to 2000 chars of free text — it IS stored (admin-only, never returned to any caller, never shown publicly verbatim) and enriched by an EU-hosted model into the agent-source day/week/month summary, kept strictly separate from the human-feedback summary. Same privacy handling as the human free text. Optional feedback (INT-1882, accepted only when the platform feedback-loop is enabled): `outcome` (correct|wrong|partial) is the minimal always-useful signal; `findings` is the rich agent signal — the real/false split per severity bucket {high,medium,low}:{real,false}. The auto-scoring already counts the buckets; you supply only whether each was a TRUE catch or a FALSE positive. Sharing the full findings earns the review-discount once go-live (one model-call less on that round). Per-account dedup: one authoritative rating per request_id per account; re-submitting updates it (last-write-wins). Requires the same API key that made the original call. Feature-gated on the platform feedback-loop: when it is disabled the tool returns a clear "not enabled" message rather than an error.

Relay the HUMAN end-user's verdict on a consensus call — the human channel (source_type='human_via_agent'), distinct from tokonomix_rate_consensus (which is the AGENT's own rating). Use this ONLY when a real human in the session actually told you what they thought of the council's answer — NEVER fabricate or guess a human answer on their behalf; if no human weighed in, do not call this tool. The `request_id` is returned by tokonomix_consensus_ask in the billing breakdown line (` · request_id: ...`) and in the `x_council.request_id` metadata field. The `choice` legend (ask the human, then map their answer): 1 = caught a blind spot · 2 = an important improvement · 3 = confirmed my approach · 4 = added nothing · 5 = was wrong/misleading. Optional: `free_text` accepts up to 2000 chars of the human's own words (trimmed and capped server-side); same privacy handling as tokonomix_rate_consensus's note — stored admin-only, never returned verbatim to any caller. Human feedback earns the (higher) human reward rate, versus the agent-source rate on tokonomix_rate_consensus. If the platform has human-feedback relay disabled, the tool returns a clear "not enabled" message instead of an error.

Single-model passthrough call. Cheaper than consensus — use for routine reasoning, tool-orchestration, classification. Returns the model's plain answer with markup billing on top. Tip: use `tokonomix_consensus_ask` instead when correctness matters.

Return a cheap version fingerprint of the canonical Tokonomix SKILL.md (the doc that tells you when to use the other tokonomix_* tools). No network call. Returns {version, sha256, last_changed, bytes}. Use this to detect that your local cached SKILL.md is stale — if your cached version differs from the returned one, call tokonomix_get_skill to refresh.

Stage large context (over the inline cap) for a grounded consensus call (INT-1817 B08). Returns an ephemeral, region-pinned upload session: a `session` id + opaque `handles`. Pass `context:{session, handles}` to tokonomix_consensus_ask so all proposers + judges read the ONE shared context-pack (build-once, in-region digest). The staged content is ephemeral (auto-purged after a short retention) and region-pinned (EU by default). NEVER pass a URL — only file contents; the server never fetches a caller URL (no SSRF). Feature-gated on the account's context-upload capability: when it is not enabled the tool returns a clear "not enabled" message rather than an error. Use inline context.inline for small payloads either way.