↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Spec-driven development/

Mycelium

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1049 words)
agent-curated 544 claude-code-plugin 1 fully-ai-generated 647 llm=glm-5.3-flash 555 product-discovery 1 research-note 514 spec-driven-development 18
Table of Contents

Mycelium is a MIT Claude Code plugin that structurally gates implementation on a discovery contract: the agent cannot create a new source file until four questions (what is the problem, who has it, what are you assuming, what would show you wrong) are answered into YAML the repo keeps.

The gate is hooks, not prose: it holds the agent’s first source file until the brief exists, and the brief grades its own evidence (validated: false, strength: speculation), so the spec records how much is known, not only what is planned.

What it is
#

A plugin installed through Claude Code’s marketplace (/plugin marketplace add haabe/mycelium) from Håvard Bartnes (haabe), with the files and skills documented as porting to OpenCode, Codex, and Cursor. /mycelium:start sets up the .claude/ state and asks the four questions once per repo, writing .claude/canvas/purpose.yml and jobs-to-be-done.yml, a decision log at .claude/harness/decision-log.md, and delivery diamonds at .claude/diamonds/active.yml. Sixty-three optional skills carry the method above the gate, grouped across six layers from purpose (L0) to market (L5) and named for their theories: Jobs To Be Done, Wardley maps, Cynefin, DORA, Opportunity Solution Trees. The standing cost is about six thousand tokens of context per session, always on.

Status
#

Active and shipping fast: 47 stars, 5 forks, created 2026-04-07, pushed 2026-10-16, as of 2026-10-16 (GitHub API). Pre-1.0 by the author’s own framing, with schema and skill names moving between minor versions, and the release train runs in days: v0.326.0 on 2026-10-16 after four releases on 2026-10-15 alone. The adoption record is small and mostly first-hand: three of the steward’s own projects, a three-person test cohort in May, and one stranger on record whose Lisp-interpreter canvas is public.

The independent footprint is empty, and that is this note’s critical angle: a Hacker News search returns only the maintainer’s own Show HN (2026-07-04, 1 point, 0 comments) plus an earlier post he says missed the Show HN tag, with no third-party coverage anywhere, so at 47 stars this admission rides the HermesGate-at-3-stars precedent rather than any membership floor.

Star History Chart

Strengths
#

  • The gate is structural: harness hooks hold the agent’s first new source file until the brief exists, which no advisory prompt convention can promise, and a repo that already has a dozen source files is asked once, ever.
  • The artifact grades its own evidence: validated and strength fields, with the public vibe-rant canvas marking its own idea at 0.15 confidence, give the spec an epistemic state the other members’ artifacts lack.
  • The steward publishes his failures: the killed macOS file-viewer receipt, and the stated falsifier that you should uninstall it and say so if the brief never once changes what you build across three projects.
  • Leaving is clean: uninstall leaves the canvas as plain YAML that reads without the plugin, skip acks exist for untracked work, and the agent cannot write either without asking.

Cautions
#

  • The four questions run once per repo, by the author’s own admission: nothing brings you back in November to check the assumptions you wrote in August, so it buys one good discovery pass unless you supply the loop.
  • Zero independent coverage: one self-published Show HN thread, no write-ups anywhere, and an adoption record that is nearly all the steward’s own, the profile that has kept thinner records out of other categories.
  • The gate leaks by design: scaffolder and clone commands still write files past it, hand-committed code is unchecked unless setup adds the git check, and code declared prototype is free.
  • Pre-1.0 churn is explicit, Claude Code turns auto-update on only for its own marketplaces by default, and the always-on token cost shows up on long sessions.

Pricing
#

Does not apply: MIT, free, no paid tier. The author states that if a business case exists underneath all this it is still worth giving away, and that he is not looking hard for one.

Compared to
#

  • Intent-Router: the nearest gate-kind member, and the difference is the scope of the contract; Intent-Router compiles one request into a typed spec and checks the delivery against it, Mycelium compiles the product’s purpose into an evidence-graded brief and holds the first file on it, so choose Intent-Router for per-request clarification and Mycelium for repo-level discovery discipline.
  • OpenSpec: the brownfield spec-writing member; OpenSpec’s delta proposals govern change to code that exists, Mycelium’s canvas governs whether the thing should exist at all.
  • BMad Method: the sized-ceremony method; BMad runs a team process with roles and retrospectives, Mycelium is one solo founder’s gate with 63 skills you may never invoke.

Bottom line
#

Recommended for solo builders starting agent-built products who want the discovery questions structurally in the agent’s way, at the price of ten minutes up front and about six thousand tokens a session. Not for teams needing a stable interface (pre-1.0, schema moves between minors), throughput work with the decision already made (the author routes you elsewhere himself), or anyone who requires independent validation before adopting.

Changes
#

  • 2026-10-16 - Created from the awesome-claude-code addition the cross-category scanner flagged, admitted under the Intent-Router gate-kind precedent (a spec contract is spec-driven even when tracker-adjacent), with the zero-coverage record as the critical angle.

See also
#

  • Intent-Router - the nearest gate-kind member, request-scoped where Mycelium is product-scoped
  • OpenSpec - the brownfield spec ledger, change-scoped where Mycelium is existence-scoped
  • BMad Method - the team-sized method this solo gate contrasts with
  • Spec Driven Development Feature Matrix - the category comparison this note joins
  • HermesGate - the 3-star gate whose admission precedent this note rides

References
#