Scaffolde is a personal AI capability factory: skills, agents, workflows, and verification kept as versioned source, then projected into Claude Code, Codex, Hermes, OpenClaw, and the other runtimes where AI operators work. Upstream infrastructure, not another chat client.
See the factory model end to end: canonical source, typed manifests, one-way projection into 12 declared surfaces, and the evidence loop that decides what lands back in source.
Explore the product →The thesis: capability as versioned infrastructure, a working single-operator system, the founder, and the path to productizing it.
See the thesis →The repository is the factory; runtime surfaces are projections. A hand edit under ~/.claude or ~/.codex is drift, not authority: legitimate live fixes get promoted back into canonical source, then re-projected.
Four steps turn a capability into versioned infrastructure every declared surface can use.
Skills, agents, workflows, and adapters are written into one versioned repository. The repository is the factory, so nothing exists only inside a runtime directory.
Manifests declare what each surface receives, and validators check the references against the file set. Missing or inconsistent references stop projection.
The sync engine writes managed output into declared runtime roots, snapshotting file hashes so it halts before clobbering a hand-edited live file.
Surfaces report ready, degraded, attention, or blocked. Evidence from the runtime decides whether a live fix is promoted back into canonical source.
Scaffolde runs on Algorithm 8.4.0, a conjecture-and-refutation loop: the goal survives verbatim, done is a set of falsifiable claims, and each claim names the probe that could disprove it. “Should work” is forbidden.
Everything below is versioned source in one repository, projected per-surface through typed manifests.
177 active skills across 26 declared domains, authored once as canonical source and projected wherever they are declared.
Nearly 40 named agents, including cross-vendor wrappers: Claude models plus Gemini, Grok, Qwen, DeepSeek, Kimi and more. Wrappers are agents in the factory, not verification coverage.
235 multi-step workflows, versioned with the capabilities they call, so a change to either is a reviewable diff.
12 declared surfaces, six of them primary executors. A declaration says what the factory can project, not that a given machine is healthy.
17 declared services: voice, browser, observability, WhatsApp, knowledge, and more, each with its own contract in source.
51 declared jobs: nightly ship, inbox checkers, memory health, secret scanning. Scheduled work is source too, not cron folklore.
Paperclip owns the durable backlog, gbrain synthesizes cross-surface knowledge with lineage, and human-authored intent wins conflicts.
The VerificationGate hook, drift detection, doctor and readiness reporting, and 34 gate validators whose registry is asserted against the validator file set in CI, so a new validator fails the build until it is registered.
Paperclip and gbrain are vendored upstreams with explicit patch boundaries. The factory carries them, it did not author them.
Counts come from the repo’s generated indexes and manifests, August 2026. Surfaces are declared, which says what the factory can project, not that every machine is healthy.
Operators now run several agent runtimes at once, and each one keeps its own copy of the same capabilities. The work compounds in the runtime, where it cannot be versioned, and never in one place that outlives the tool.
Claude Code, Codex, Cursor, and a wave of agent CLIs each keep their own skills, prompts, and settings under their own root directory.
The same skill gets rewritten per tool, drifts per machine, and dies with the session that produced it.
A live tweak under ~/.claude is unversioned, unreviewed, and silently lost the next time the tool updates its own files.
A canonical factory that versions capability once and projects it everywhere, with drift detection holding the line and evidence deciding what returns to source.
Models and agent CLIs turn over every few months. What an operator has actually built, the skills, agents, and workflows, should not turn over with them. Scaffolde keeps that layer as versioned source and projects it into whatever runtime comes next.
Capabilities kept as versioned source outlive any single model, vendor, or CLI. Scaffolde sits upstream of all of them: skills, agents, workflows, and verification machinery in one canonical repository.
Typed manifests plus a one-way sync engine put one capability catalog onto 12 declared surfaces. Drift detection keeps the repository authoritative: a hand edit under a runtime root is drift, not authority.
Evidence-backed completion is the practice that keeps a compounding system trustworthy to its operator: falsifiable claims written before the work, fail-closed gates, and 10,900+ maintainer-health tests green.
The status, stated plainly: private, single-operator, macOS, with v2.0.0 as a release candidate. Everything below is checkable in the repository.
Scaffolde is the factory. Runtime surfaces are projections.
No form, no sales call. Read how canonical source becomes a projected surface, and what evidence it takes to change the factory.