REFERENCE
Know what MOMM does.
And what it cannot prove.
The protocol is intentionally stricter than a model’s confidence. These are the boundaries to understand before using it on important work.
Who does what?
- Governor
- Your current coding agent. It remains the sole writer, reproduces material findings, verifies changes and records decisions. It is excluded from peer review.
- Reviewer route
- An installed provider CLI using an OAuth/account session. Its inner model depends on your account and configuration; a CLI name is not a model ID.
- Finding
- A reviewer's claim about a defect. Agreement affects investigation priority, not truth.
- Disposition
- The governor’s recorded ruling: applied, applied-with-modification, rejected or deferred. Applied rulings name checkable verification.
Not just code: prose and supported attachments
Use MOMM for manuscripts, specifications and other text when sharing with the selected providers is permitted. The same rule applies: reviewers make claims; the governor verifies and records decisions.
Text routes include Codex, Claude Code, Antigravity, Copilot and Grok. Verified attachment adapters differ: Codex supports images; Claude supports images and PDFs; Gemini supports images, PDFs, audio and video where the account is eligible. Other routes stay text-only until verified. Run preflight for the installed adapter’s actual capability; a provider logo is not evidence of multimedia support.
The historical public ledger contains the manuscript specimen rev_20260904131823_wvxh. Inspect the sanitized specimen → · Read the current protocol ↗
Strict review-contract rejection and input/source size ceilings remain possible. The source completion validator covers local text and supported Git text additions/modifications, not every binary, rename, deletion or media lifecycle.
What leaves your machine?
Review input, optional project rules and the review contract go to the selected providers through their official CLIs. Inputs are sanitized first, but no automatic redactor can certify that a document is safe to share.
Prompt instructions are not filesystem isolation. Provider CLI tools may discover additional local metadata; use approved source-sharing contexts and platform permission controls. On Windows, ordinary helpers surviving a normal leader exit are not guaranteed to be contained; stronger OS job ownership remains open.
The daily version notice requests a public manifest. Setup Center’s explicit maintenance checks may query package versions and provider model lists. Connectivity tests send a disclosed synthetic sentence. Review history is not uploaded by these checks.
The private evidence directory stays local and gitignored. Publishing is a separate, explicit sanitized-export workflow. Never publish another user’s ledger by treating it as a page asset.
Start with the actual failure
| Status | Meaning and next action |
|---|---|
| authentication_required | Use the exact login_hint emitted for that provider. Complete its browser login yourself. No API-key fallback. |
| provider_unavailable | A transient provider outage; MOMM already retries once. Wait and retry later. Re-login is not the default fix. |
| ineligible_tier | The provider rejected that account tier. Read the reported licensing/migration guidance; another login will not change eligibility. |
| missing / unsupported | The CLI is absent or the requested capability has no verified adapter. Read preflight’s specific explanation. |
| timeout / invalid_output | No usable review was obtained. Inspect the detail, narrow the input or choose another ready route. Do not count this as an ACCEPT. |
| self_excluded | The governor is correctly excluded. It is not a failed external review. |
| Update verification refused | Read the named prerequisite, signature, hash, dirty-tree or receipt error. Do not bypass it with a manual pull and call the update verified. |
Useful commands
node momm/scripts/multi-review.mjs --help
node momm/scripts/multi-review.mjs --doctor --pretty
node momm/scripts/multi-review.mjs --stats
node momm/scripts/multi-review.mjs update --helpFor a project review use the installed script’s absolute path while staying in that project. --min-success 1 requires at least one successful external reviewer; use an appropriate higher quorum for a release gate. --stream keeps structured progress separate from the final report.
Does it mechanically enforce every reproduction?
No. The governor validator checks source hashes, item decisions and recorded test evidence; it cannot prove that an observation is honest or a test adequate. It never executes peer-supplied test instructions. Treat the tests, code and disposition evidence as the checkable record.
Is a unanimous ACCEPT enough?
No. It only says the successful reviewers returned ACCEPT. Missing, timed-out or invalid routes do not join that verdict, and agreement never replaces your project’s tests.
Does it work identically on every machine?
No blanket guarantee is justified. The CI matrix is configured for Windows, macOS and Linux on Node 18, 20, 22 and 24, plus pinned Windows builds on Node 24.15.0 and 24.19.0; check its result for the exact revision you use. Actual provider behavior also depends on CLI versions, account eligibility, quotas, network access and sandbox access to OAuth stores.
What remains open?
Large-input route reliability, containment of independently detached processes, unsupported source-snapshot types and broader independent evaluations remain distinct work. This release does not claim to have solved them with an update command or a new website.
Common questions about MOMM
What is MOMM?
MOMM (Mixture of Model Modality) is an open-source Agent Skill for multi-model code and document review through installed AI CLIs. Your current agent stays the sole writer and verifier; other reviewers return claims, and your agent records the decisions in a private local ledger.
Which AI reviewers can MOMM use?
The default reviewer pool is Codex, Claude Code, Antigravity, GitHub Copilot and Grok. Gemini is opt-in for eligible accounts. MOMM excludes the current governor and uses available, authenticated routes; these CLI names do not prove which underlying model answered.
Is MOMM entirely local, and is my code private?
Orchestration and the private review ledger are local. Selected providers receive the sanitized review input through their CLIs. Redaction is not a confidentiality guarantee: approve the provider set and sharing scope before using private material.
Is MOMM free, and does it require API keys?
MOMM is MIT-licensed open-source software. Provider subscriptions, usage limits and quotas still apply, so a review is not necessarily free. MOMM uses existing OAuth or account logins and has no API-key fallback.
How do I install or upgrade MOMM?
Use the install or upgrade prompt to identify your harness and any existing installation, verify a published stable release, preview the changes and ask before applying them. Keep the clone in a permanent folder. Protocol changes require separate acceptance; automatic updates are off by default and require owner opt-in.
Where are the dashboard and review results?
Setup Center manages reviewer connections and CLI versions. Each reviewed project has its own local ledger containing reports and governor decisions. This public website provides guidance and a separately sanitized historical example; it is not your private dashboard.
Does reviewer agreement prove that code is correct?
No. Findings are claims, and a unanimous ACCEPT is not a test. The governor investigates findings, verifies changes and records every decision. The published development snapshot is evidence of this project’s workflow, not an independent accuracy benchmark.