Audience: prospective users + evaluators with the questions Mycelium has been asked. Time to read: 5 min. Last updated: 2026-05-08.
The first six questions surfaced at the 2026-05-07 Juniors.dev presentation; the rest are standard OSS.
Honest answer: depends on session shape. Mycelium itself adds ~6,000 tokens of harness overhead per session (hooks, gates, pre-task context). The bigger cost is what the framework refuses to skip — /interview is 15-30 minutes of structured Q&A that an unharnessed agent would skip. That cost is borne whether it's worth it or not.
External vendor estimate (2026-04): Claude Code averages ~$13/day for an active developer. Mycelium's overhead is well under 10% of that on the projects observed so far. The framework is configuration files, not a service — there is no Mycelium subscription.
If token cost is the binding constraint and discovery is settled, the framework is over-engineered for the work. See Who it's not for.
No collision in Claude Code's resolution order: user-defined skills take precedence over Mycelium's. If you have /interview already, yours runs. Mycelium discovers skills from .claude/skills/*/SKILL.md and registers them as candidates; it does not override existing names.
If you want Mycelium's /interview AND your own, rename one. The framework has no opinion on which.
The framework's value concentrates upstream. If your problem statement is settled, your user is named, your scope is bounded, and your only friction is execution velocity — Mycelium adds ceremony with thin marginal value. Use Addy Osmani's agent-skills or paddo.dev's boring agents patterns.
If your idea is rough — you're not sure who it's for, or whether it's a real problem, or what "done" looks like — Mycelium's discovery scales (L0–L2) earn their cost. The clearer the idea, the less you need; the rougher the idea, the more you do.
The same heuristic stated as a rule of thumb: the rougher the idea, the more scales earn their cost.
Partially. Mycelium's full enforcement layer (hooks, gates, reflexion loops, framework-guard, secret detection) is Claude-Code-specific today. The portable surface is canvas reading + writing + decision logging — non-Claude-Code agents can read .claude/canvas/*.yml and .claude/memory/corrections.md to operate on the same evidence base, but gates do not fire automatically.
See AGENTS.md for the minimal path on non-Claude agents and concrete operating models per agent class.
Using a tool on its own development. Mycelium's framework is dogfooded on Mycelium itself — the friction the founder hits while building Mycelium becomes corrections that shape Mycelium.
In Mycelium specifically, the meta_dogfood project type formalizes this: when a project's purpose is "improve the framework", canvas writes target the framework's own discovery scales rather than a downstream product's. See philosophy.md for why this is required, not optional.
Canvas files are YAML committed to git. Conflicts resolve through git's normal merge — same as code. Mycelium's specific discipline: _meta blocks (version, last_validated, evidence_type, structural_level) are the source-of-truth markers; if two branches edit different fields of the same canvas section, the merge is mechanical. If two branches edit the same field, the higher last_validated date wins by convention (the more recent evidence supersedes the older).
/canvas-sync packages canvas state for cross-session sync; it does not auto-merge — git does that. See usage-modes.md for the team workflow.
MIT License. Commercial use is permitted. Mycelium is not affiliated with any of the cited authors or publishers — citations are educational and credit the intellectual foundations.
See CONTRIBUTORS.md for who's listed and the "How to get listed" recruitment-shaped section. See docs/contributing/ for process.
The framework's bias is to credit the named person who shaped a mechanism, not the maintainer who merged it. Surface friction; if it shapes the framework, you land in docs/receipts/.
Every case file in docs/receipts/cases/ carries YAML frontmatter naming the commit hashes the case produced. Click through to the commits on GitHub. The receipts argument stands on its receipts.
No — Mycelium is configuration files plus orchestrated prompts, not an AI system in the Act's definition. Products built with Mycelium may be regulated. See regulatory.md for the mapping table and /regulatory-review for the per-product check.
Mycelium does not certify compliance. For compliance decisions, consult qualified EU AI law counsel.
/interview routes on canvas state, not on a time budget. An empty canvas gets the universal brief; a populated one resumes via /diamond-assess. Depth and time-cost are chosen after the brief, when you have the data to choose. Project complexity is then inferred from canvas density and the number of L2 opportunities — more opportunities and more scenarios mean more diamond depth.
(Corrected 2026-07-30: this answer described a Phase 0 time-budget selector — under 8h, 8–48h, 48h+ — that was removed. It asked the reader to predict the future before any value had been delivered, which is why it went.)
You can override at any time. The framework recommends; it does not force.