Changelog
API field additions, receipt version bumps, and docs semantic changes.
September 2026
Buyer API
- Sticky conversations are live. Send
X-MacProvider-Conversation: <tag>to prefer the provider that served earlier turns, for up to 1,800 seconds. It is a preference, not a guarantee, and often routes elsewhere; only sticky hits bill at the cache-hit rate.DELETE /v1/stickyclears your sticky state. The live posture is intier1_disclosure.sticky_affinityonGET /v1/models - Automatic prefix-cache reuse for OpenAI-compatible clients: repeated system prompts and tool scaffolds can reuse a Mac’s KV cache without any header. These hits are reported in the OpenAI-standard
usage.prompt_tokens_details.cached_tokensand billed at the full prompt rate - Tool calling: tool-call IDs from other OpenAI-compatible clients are accepted and rewritten;
tool_choice: "required"or a named function is served asauto; streamed tool calls are safe to concatenate - Next provider release (not in stable v1.8.123): removal of the 256-message cap on chat requests (stable providers still return
400 messages_too_longabove 256 messages), and recovery of incomplete Qwen XML tool-call wrappers and leftover tool markup instead of failing the stream - Routing skips providers below a minimum throughput floor, and seat accounting no longer hides free seats when a Mac reports “busy” late
- Settlement robustness: a served completion is returned to the buyer even if recording its settlement evidence hits a database deadline, and settlement holds are reconciled promptly
- New reference pages for the Beta OpenAI Responses facade (
POST /v1/responses) and the experimental Anthropic Messages facade (POST /v1/messages)
Models and pricing
- The signed model catalog was re-signed on 23 Sep 2026 to cover 17 models, and it is live. Stable provider v1.8.123 fetches the live 17-row feed and falls back to its compiled-in 10 rows if the fetch, signature, or age check fails. See Model catalog
- Seven new priced rate-card rows: Qwen3 30B A3B Instruct 2507, GLM-4.5-Air, Qwen3.5 27B, Qwen3.5 35B A3B, Qwen3.6 27B, Qwen3.6 35B A3B and Qwen3.8 27B
- gpt-oss-120b joined the autotune catalog for 90 GB+ Macs
- The rate card is now a signed feed derived from a dated market snapshot, with a guard that stops the signed rates and the billing configuration from drifting apart
- Corrected the Qwen3.6 27B catalog artifact identity so providers serving it match the signed row. Parts of this fix are provider-side and ship in the next provider release
Providers
- Provider CLI v1.8.123 (8 Sep 2026, GitHub Latest):
malibu-clicommand (a symlink to the canonicalmacprovider-cli, which keeps working)malibu-cli doctor report --jsonread-only diagnostics report and one consistent provider statusmalibu-cli autotune --calibrate-contextto size context after model selection- Model-recommendation adoption in the Malibu app (Adopt / Not now)
- Verified model weights kept in a durable app-owned store
- Installer and updater recovery, including
malibu-cli recover-updatefor a wedged update - Idempotent uninstall that keeps saved provider access
- Autoupdate reliability: recovery through the signed release channel when a session is stuck, and backoff on repeated failures
- Fresh-install fixes for hangs and for Macs where no paid model fit
- Installer fixes served from
get.malibu.techafter v1.8.123: latest-release lookup no longer fails with a 404; fresh Macs install without the Command Line Toolspython3stub; no falserollback_failedon a first install - Headless installs over SSH for Macs without a logged-in GUI session
Docs catch-up (late August)
- Public, no-key endpoints on
api.malibu.tech:/v1/rate-card,/v1/stats/overview, and the routability views/v1/stats/models,/v1/stats/routabilityand/v1/stats/providers GET /v1/receipts/{request_id}: metadata-only receipt status for the requesting account- Malibu dashboard: payout wallet registration (it does not send funds) and a preview off-chain reward ledger; earnings open through a portal link
- Apple hardware attestation reported for a first provider; no routing or settlement effect yet
Documentation
A full refresh against the shipped system on 23 Sep 2026. Corrections:- Header names. There are no
X-Malibu-*headers. Real ones includeX-MacProvider-ConversationandIdempotency-Key(request) andX-Provider-Id,X-Request-ID,X-RateLimit-*,Retry-After,X-MacProvider-ReceiptandX-MacProvider-Streaming-Mode(response). See Headers for the full list - Receipt verifier. The tool is
macprovider-verify(releaseverify-v1.1.0);malibu-verifynever existed. Offline verification of current receipts is pending a verifier update: 1.1.0 cannot parse today’s v0.3 buyer receipts. Settlement receipts (v0.4) are checked by the network and are never sent to buyers. Some non-streaming responses carry a v0.3 receipt header (currently roughly a quarter to a third); streams never do. For the rest, use/v1/receipts - Provider CLI. Config lives at
~/.config/macprovider/config.yaml, environment variables use theMACPROVIDER_prefix, and the LaunchAgent label islive.malibu.provider. Invented flags, commands, env vars and~/.malibupaths were removed - Autoupdate. Releases are checked with an ECDSA P-256 signature over
checksums.txt, not Ed25519. Nonexistentchannel,drain_timeout_sandautoupdate rollbackoptions were removed; manual recovery isrecover-update - Structured output is validated after generation, not enforced by constrained decoding
- Pinning. Buyers cannot pin a request to a provider; sticky conversations are the supported way to prefer one
- Cache billing. Only sticky-conversation hits bill at the cache-hit rate; automatic prefix-cache hits bill at the full prompt rate
- Settlement. Outcomes are
verified,pending,quarantinedandzero_settled, with a 300-second pending deadline - Model catalog. The 17-model signed catalog with per-model memory floors replaces older lists. Qwen2.5 7B and Llama 3.3 70B are not in the catalog
- Money. Public buyers are not charged in pre-beta (daily token quota); USDC payouts are off; wallet registration only records an address
- Routing tiers. New Macs move from provisional to trusted by meeting trust criteria; pinned is set only in operator configuration
- Litepaper. Removed per-Mac annual dollar figures and hardware-payback lines, and labeled the PoMW prototype numbers as unaudited
- New pages: Trust disclosure, Responses, Messages
August 2026
Product (docs catch-up)
- Agent-readable onboarding at
https://get.malibu.tech/skill.md— a coding agent can install the provider CLI, check status, recover a broken setup, update or uninstall, and connect apps toapi.malibu.tech. Discovery index:https://get.malibu.tech/.well-known/skills/index.json - Public snapshot is
GET https://api.malibu.tech/v1/stats/overview(also onmalibu.tech).GET /v1/network-statsis a compatibility alias for the same body - Public rate card is
GET https://api.malibu.tech/v1/rate-card(also onmalibu.tech) GET /v1/modelsrequires a Bearer token- Provider admission is invite-gated pre-beta (malibu.tech/host); fresh register needs
malibu.tech/j#/… - USDC-on-Base payout rail marked Planned (beta) on Network status — credits accrue in pre-beta; USDC payout ships when Malibu enters public beta
- Download button follows GitHub Latest and links the immutable GitHub release DMG for that tag, never a moving
latest.dmg - Console: agent tools, stop control, activity bus, client-measured decode tok/s on the bill line
- Catalog vs warm set documented: priced catalog rows route only when a matching-RAM Mac is serving them. Warm snapshot on 16 Aug 2026 was 3B + Qwen3-8B (the model list in this entry was corrected in September 2026)
- Homepage and
/hostearn copy matches the ledger: 90/10 credits live in pre-beta; USDC payout at public beta; $MALIBU / 70/12/18 planned - Operator named on site: Malibu, jurisdiction Cayman Islands (privacy, terms)
Documentation
- Ledger, catalog, stats, and example model IDs refreshed against live APIs (16 Aug 2026)
- Added Economics; pricing page now has a dated OpenRouter snapshot
- Buyer / provider / network nav groups
July 2026 (continued)
Documentation — claim validation (Phase B)
- Added evidence pages: Benchmarks & methodology, Pricing comparison, Provider economics, Threat model, TOPLOC integration, Glossary
- Litepaper quantitative claims now link to evidence pages instead of standing alone
- CI:
scripts/validate-docs-claims.mjschecks Litepaper ↔ evidence cross-links
July 2026
Documentation
- Mintlify docs source in
MalibuAI/malibuunder/docs; self-hosted on Vercel vianpm run docs:build(no Mintlify GitHub org connection required). - Litepaper remains
/docsdefault opening page - Added Network status, Roadmap, and Changelog ledger pages
- Litepaper updated with live vs planned status labels and trust-boundary pointers
API (live)
- Receipt tuple v0.3: nine-field signed buyer receipt (September 2026 correction: v0.3 receipts are not settlement receipts; settlement uses v0.4)
usage.cached_prompt_tokens— always present; non-zero only on a sticky hit