Myzelium turns meetings, decisions, specs, code, and tests into one living graph — source-backed context your team and its agents build from.
Access is invitation-only today. Joining the waitlist puts your team in line for the next cohort.
We benchmarked Myzelium's spec graph against four file-based spec-driven development frameworks — GitHub Spec Kit, BMAD-METHOD, GSD, and Kiro — on real change requests in two production codebases, with the same model driving every run.
| Per change request | File-based SDD | Myzelium |
|---|---|---|
| Where specs live | Markdown files scattered through the repo | A typed graph — states, owners, relationships |
| “What does this change affect?” | grep, then read enough code to be sure | One impact query over typed edges |
| Rules that aren’t in the code | A dead end — nothing to find | Retrieved from the spec that states them |
| Prior work with acceptance criteria | Reconstructed from source, when it exists | Returned as linked sibling specs |
| Cost as the product grows | Grows with the codebase | Flat — one or two queries per change |
What runs today. Reading source catches drift, dead wiring, and unfinished substrate — file-based tools are genuinely good at this.
Rules, decisions, and acceptance criteria that may not be in the code at all. In our benchmark, one governing rule existed only in the spec graph — no amount of code search could surface it.
Myzelium keeps both connected: agents read your code and query the graph, so intent and reality stay in the same conversation.
Internal benchmark, July 2026 — four change-request scenarios, two production codebases, the same model driving every framework in isolated worktrees. Kiro modeled, not executed. Full methodology available on request.
Every spec-driven framework we surveyed stores specs the same way: flat markdown files in a repo. Myzelium promotes them to typed nodes in one persistent graph — and that single decision is where the benchmark numbers come from.
One query, a typed-edge traversal, a composed brief — the round trip an agent makes before every change.
Contains, relates to, depends on — relationships are data you can query, not prose you have to read.
Draft, approved, implementing, tested — every spec knows where it stands, and drift gets flagged.
“What depends on this?” is one call over the graph — the blast radius of a change, before you make it.
Agents receive a few thousand characters of exactly-relevant context — sources, criteria, prior art — never whole documents.
Swipe to compare all five
| Capability | Spec Kit | BMAD | GSD | Kiro | Myzelium |
|---|---|---|---|---|---|
| Specs as a typed graph, not flat files | Not documented | Not documented | Not documented | Not documented | Yes |
| “What depends on this?” as a single query | Not documented | Not documented | Not documented | Not documented | Yes |
| One persistent graph across the whole product | Not documented | Not documented | Not documented | Not documented | Yes |
| Spec health — drift, orphans, lifecycle states | Not documented | Partial | Not documented | Not documented | Yes |
| Traceability from spec to code, commit, and test | Not documented | Not documented | Not documented | Not documented | Yes |
| Live retrieval over MCP | Not documented | Not documented | Not documented | Yes | Yes |
The trade, honestly: flat files are zero-setup and review like code in a pull request. The graph is a running service — heavier to adopt, and it pays back as the product and the number of agents touching it grow.
Marks reflect each framework's primary documentation, July 2026 — a dash means not documented, not impossible. Kiro's MCP retrieval is genuine prior art, and we say so.
Meetings, docs, decisions, code, and tests become source-backed specs — each one knowing its sources, state, and downstream effects.
Clarity drafts, flags contradictions, and proposes work. Nothing becomes the source of truth until a person approves it.
Coding agents receive composed briefs — impact, acceptance criteria, prior art — and ship changes traceable back to their specs.
We're onboarding teams in small cohorts. Tell us where to reach you.