Your agents cite
the record, not vibes.
MsgStack keeps a locked, versioned record of what your company is allowed to claim — compliance rules, product specs, brand positioning — and makes every AI agent cite it.
Locked claims return verbatim, never paraphrased.
Vector RAG converts your policies into float arrays and approximates answers — under pressure, models paraphrase compliance rules and invent metrics. The canon graph traverses exact boundaries instead: when a root policy changes, every downstream artifact flags for review before anything drifts.
003 — See the audit trail ✦> get_entry 441-B "98% of zero-day threats stopped proactively." LOCKED ✦ · v12 · Product Marketing 7 artifacts bound · drift 0%
A canon layer, not a marketing tool.
Every department gets a grounding type: a schema that structures its knowledge into queryable, governable canon. Message House is the wedge — Engineering Spec and Policy Shield are the proof.
Message House
Product Marketing
Brand pillars, proof points, personas, objection handlers. Every generated claim cites an approved entry.
Engineering Spec
Engineering
API constraints, SLAs, deprecation notices. Copilots aligned with real specs — not hallucinated rate limits.
Policy Shield
Legal & Compliance
Disclaimers, privacy rules, compliance assertions. Legal language retrieved verbatim — never paraphrased.
Canon without governance is suggestion.
Five lifecycle states gate what grounds. Three content tiers define how strictly each entry must be reproduced — Tier 1 is verbatim, enforced. Every claim has a named DRI, and every grounding query lands in an exportable audit log.
The governance layer ✦Draft → In Review → Approved → Locked → Outdated
Verbatim · Structured · Grounded
Named, accountable, transferable
Every query on the record
A real session, end to end.
Discover
list_canon_domains returns every approved truth framework, with entry counts and lock status.
Retrieve
search_canon pulls version-locked claims. Exact matches, never paraphrased.
Generate
generate_artifact produces output cited directly against the active domain.
Audit
score_alignment flags any claim without a matching locked entry before it ships.
Deterministic. Locked claims return verbatim — the graph traverses, it never guesses.
Hard contradictions caught by the multi-agent audit before anything goes out.
Local SQLite verification on every query. Self-hosted, no round trips.
Run the checker yourself.
MsgStack validates AI output against approved canon assertions. Simulate scanning a generated SDR email for ungrounded claims.
Verification Audit
Self-hosted, yours entirely.
The core product — MCP server, admin UI, knowledge graph, artifact generation — is and stays open source. Local vector index, zero external dependencies, one docker compose.
Request Early Access
MsgStack is open source and self-hostable. Join the developer waitlist for early access to updates, new grounding schemas, and MCP tooling.