Real tool call → trusted evidence
When an MCP proxy sees client↔server frames, turn raw JSON-RPC into a redacted receipt proving capability agreement, call status, replay evidence, and privacy defaults.
Pluribus reviewer shortcut
Context receipts, grouped by the boundary they prove.
Most AI-agent tools now promise memory, context, skills, MCP gateways, or multi-session orchestration. Pluribus is narrower: it gives reviewers small, privacy-safe receipts for the moment context crosses into authority.
No raw prompts, source files, transcripts, customer data, tool outputs, secrets, or full schemas are required for these demos.
Untrusted package → reviewable behavior
When a sandbox captures npm/PyPI/MCP behavior, turn raw traces into a model-safe receipt proving target hash, sandbox policy, artifact hashes, behavior counts, and verdict.
Repo memory → allowed authority
When live memory, RAG, or a knowledge graph answers a repo question, prove snapshot freshness, cited refs, omitted private data, authority level, and stale-if rules before edits.
Claim → evidence attestation
When a skill, registry, leaderboard, review, approval, or handoff makes a reusable claim, emit evidence refs, omissions, limits, verdict, and stale-if rules without private logs.
Skill invocation → honest portable receipt
skill-graveyard now exports a privacy-safe skill-use receipt that omits local paths, prompts, and tool output. It proves invocation without pretending to know whether the skill changed downstream work.
Commit diff → agent change manifest
When Claude Code, Codex, Cursor, or another agent writes a commit, keep Git as the byte ledger and add a privacy-safe sidecar for intent, loaded authority, checks, omissions, trailers, and stale-if rules.
Canonical rules → loaded target authority
When a config doctor merges AGENTS.md, CLAUDE.md, Cursor rules, Copilot instructions, Skills, hooks, or settings, prove each target actually loaded the treated config and record the drift gate.
Hidden inventory → model-visible tool
When Tool Search or a local gateway hides hundreds of MCP tools, prove which server/profile and upstream definition hash surfaced behind the alias.
Durable instruction → current authority
When CLAUDE.md or AGENTS.md has aged for months, require live evidence, revisit dates, and demotion rules before the agent obeys it.
Human-visible text → agent-read stream
When Unicode, bidi marks, homoglyphs, generated Skills, or inline expansions can change what the agent reads, require byte ranges and a review gate.
Extension surface → active authority map
When CLAUDE.md, output styles, Skills, hooks, subagents, plugins, and MCP can all affect one run, source-label the active layers before trusting a write.
Session state → safe next action
Before /clear, compaction, handoff, or another write, prove which files/commands support the current plan and what state was lost or kept.
Multiple sessions → operational ownership
For Claude Code/Codex/OpenClaw sessions running in parallel, review path claims, last evidence, blockers, and unsafe-to-resume state instead of trusting a chat summary.
Retrieved memory → allowed authority
When memory servers or knowledge graphs return facts, separate used, suppressed, stale, and rejected facts before they become durable project authority.
One-minute local check
Run one public demo without cloning the repository. It validates a redacted memory answer receipt for snapshot freshness, cited refs, private omissions, authority level, and stale-if rules.
npm exec --yes --package github:caioribeiroclw-pixel/pluribus -- pluribus demo memory-answer-receipt --json
npm exec --yes --package github:caioribeiroclw-pixel/pluribus -- sh -c 'pkg=$(dirname "$(readlink -f "$(command -v pluribus)")")/..; node "$pkg/examples/evidence-attestation/check-evidence-attestation.mjs" "$pkg/examples/evidence-attestation/evidence-attestation.json"'
For directory maintainers and newsletter reviewers, the Community Review Packet has copyable listing text and disposable smoke tests.