boost — a package manager for AI coding skills
Homebrew for AI coding skills. boost finds, installs, and version-tracks skills from GitHub-hosted registries, and wires each one into Claude Code, Windsurf, Cursor and Gemini CLI in a single pass.
pipx install boost-skill-cli
boost tap --defaults # 7 curated registries: skills, rules, workflows
boost search tdd # full-content BM25 over every skill
boost install tdd-workflow # → every agent, version-pinned, one lock file
No more copying SKILL.md into four agent folders and forgetting which one is
stale. The default install is pure stdlib, with no build step and no
dependencies.
tap a registry → search → install into every agent → doctor — one flow
What's a skill, exactly?
A skill is a SKILL.md file: Markdown plus a bit of YAML up top, giving an
agent a repeatable capability. A workflow to follow, a set of house rules, or
domain knowledge it wouldn't otherwise have. Agents like Claude Code pick these
up automatically out of ~/.claude/skills/.
The tedious part is everything around that file: finding good skills, wiring each one into every agent you use, keeping versions in sync, and handing them to teammates. That is what boost takes over.
Install
Needs Python 3.12+ and git.
pipx install boost-skill-cli # or: pip install boost-skill-cli
boost --version
boost tap --defaults # pull in the 7 starter registries
Upgrading later is upgrade, not install --upgrade: with a bare package name
the latter matches the spec you already satisfy and silently does nothing.
pipx upgrade boost-skill-cli # or: pip install --upgrade boost-skill-cli
Or run it from a checkout. The runtime is stdlib-only, so there is nothing to install beyond the shim:
git clone https://github.com/jonnyeclectic/boost ~/.boost-src
ln -s ~/.boost-src/boost ~/bin/boost # anywhere on PATH works
Search
boost search ranks with a built-in full-content BM25 engine. The index builds
itself on your first search, so there is nothing to set up.
BM25 matches words, which fails when your query shares no vocabulary with the skill you want. Semantic search fixes that, needs no API key, and takes two commands:
pip install "boost-skill-cli[rag]"
boost reindex --dense
When both indexes exist boost fuses the rankings rather than choosing between them. See docs/semantic-search.md for the whole setup, what it costs, and how to tell which engine is actually serving.
The registry catalog
boost tap --defaults gives you seven registries. The full catalog is 480+
classified GitHub registries of skills, Cursor/Windsurf rules and Claude Code
workflows, indexing roughly 30,000 items.
boost tap --catalog --dry-run # browse the classified catalog
boost tap --catalog --type skill --limit 20 # the 20 biggest skill packs
boost tap --catalog --category infra # Docker/Kubernetes/OpenShift packs
boost tap --catalog --category java # Spring Boot / Kotlin / JVM packs
boost tap --catalog --category security # filter by category
Every est_items is measured from the repository's file tree rather than
estimated, so --limit ranks by real size. Items are counted once however many
agents a repository renders them for, so a pack shipping a copy in .claude/,
.cursor/ and twelve other dotdirs is not credited with fourteen skills.
Seven domains are curated end to end and each is tappable on its own: ai,
architecture, ui, java, ecommerce, infra and marketing. Smaller
categories fill in around them.
One of those deserves a note. efficiency collects packs whose items exist to
make an agent emit less: less code (ponytail) or fewer output tokens
(caveman). Their focus lines describe what the items do and omit the
headline savings on purpose. Independent paired benchmarks reproduce roughly a
fifth of what each advertises, with no measured quality loss, so the effects are
real but much smaller than the marketing figure.
Share the catalogue instead of re-tapping it
Tapping is the slowest thing a new install does, and the expensive artifact is not the valuable one. On a machine with 458 registries tapped the shallow clones are 12 GB, while the catalogue they produce is 10.9 MB gzipped. So hand someone the catalogue:
boost catalog --export catalogue.tgz # 458 taps · 59,972 entries · 10.9 MB
boost catalog --show catalogue.tgz # what is in it, without importing
boost catalog --import catalogue.tgz # on the other machine
Measured on a fresh HOME: import takes 0.3 s, boost reindex a further 3.8 s,
and boost search then returns real hits over all 59,972 items with zero
repositories cloned. install still clones, but only the one registry your
chosen skill lives in.
Bundles carry catalogue JSON and nothing else. Derived indexes are excluded because they rebuild in seconds, and vectors because they are only valid inside the embedding space that made them. Import merges, so it never drops taps the receiving machine already had.
Under the hood
GitHub registries ──boost update──▶ ~/.boost/repos/ (blobless sparse clones)
~/.boost/cache/ (JSON catalogs)
──boost install─▶ ~/.agents/skills/ (canonical store)
.skill-lock.json (v3 lock file)
────symlinks───▶ ~/.claude/skills/ ~/.windsurf/skills/ ~/.cursor/skills/
A tap clone holds only the files boost indexes: SKILL.md, workflow Markdown
and rule files. Everything else a registry ships is neither downloaded nor
checked out. Shopify/agent-skills is 611 MB as a normal clone and 11 MB as a
tap, and produces an identical catalog. When a skill you install owns its own
scripts/ or assets/, those are fetched on demand at install time.
If you tapped heavily before this landed, boost compact narrows the clones you
already have, offline and with no loss of search coverage:
boost compact --dry-run # what it would reclaim
boost compact # narrow every clone in place
boost compact --reclone # also drop downloaded git objects (needs network)
boost indexes three item kinds from the same registries, and they install to
different places. A skill is copied into the canonical store and symlinked out.
A rule is materialised into each agent's context file (~/.claude/CLAUDE.md,
~/.gemini/GEMINI.md), so it has no directory to look in. A workflow is
rendered per agent.
boost list --kind rule # just the rules, and where they landed
boost list --kind workflow # and which slot each one fills
Gemini CLI is the one agent that needs no symlink. It implements the
Agent Skills standard and discovers ~/.agents/skills
directly, so linking into ~/.gemini/skills as well would put the same skill in
two of its discovery tiers. Rules and workflows still materialize under
~/.gemini/.
User scope vs project scope
The default is user scope: the skills you want everywhere. The other half is the
npm --save half, for skills a team agrees on, which belong in the repository
where they can be reviewed in a PR and arrive with a git clone.
boost install code-review --local # into THIS repo (= --scope project)
boost install code-review # into your user config (default)
boost list --local # just what this repo carries
boost uninstall code-review --local
A --local install writes real directories into the repository's own agent
config and records them in a committable per-repo lock:
<repo>/.claude/skills/<name>/ real copies, not symlinks
<repo>/.cursor/skills/<name>/
<repo>/.boost/skill-lock.json commit this — it's what teammates clone
Real copies, deliberately: a symlink into your ~/.agents/skills is a dangling
pointer on anyone else's machine. The two scopes use separate lock files, so
vendoring a skill into a repository never fights with the copy you use
everywhere, and boost install --local run from src/deep/nested walks up to
the repository root rather than scattering a .claude/ three directories down.
After a fresh clone, boost sync re-materializes anything the project lock
records but the checkout is missing. It never deletes a skill directory boost
didn't write, and won't overwrite a hand-written .claude/skills/<name>/
without --force.
boost update covers user scope only for now. To refresh a vendored skill, use
boost install <skill> --local --force.
The usual workflow
boost search jira # look across every tapped registry
boost install my-jira # copy, link, lock — with a quality score attached
boost doctor # broken links, lock drift, stale taps
# hand the whole setup to a team:
boost bundle dump > Boostfile # everyone else runs: boost bundle install
Browsing the catalogue
boost browse is a full-screen picker over every tapped item: a filterable list
on the left, the selected item's detail on the right.
╭─ ●●● boost browse ──────────────────────────────────┬──────────────────────╮
│ ❯ co re 2/4 │ code-review v1.0.0 │
│ match (●) all ( ) name ( ) descr ( ) tap │ ──────────────────── │
│ ↑↓ move ⇥ select → detail ^T scope ↵ install │ Review a pull … │
├───────────────────────────────────────────────────────┤ ● installed │
│ ★● code-review [skill] v1.0.0 [acme/quality] │ SOURCE · FRONTMATTER │
╰─ ✓ 1 installed ───────────────────────────────────────┴──────────────────────╯
| key | does |
|---|---|
| any character, including space | types into the query |
↑ ↓ |
walk the whole surface: query → match toggles → results |
→ ← |
into and out of the detail pane, where ↑/↓ scroll it |
⇥ |
select, and multi-select for a batch install |
^T |
cycle the match toggles without leaving the results |
↵ |
install, in place, without leaving the browser |
esc |
quit |
The query splits on spaces and every token must match, so code review narrows
where code alone would not. Selection lives on ⇥ rather than space for
exactly that reason, and q stays a character so quality is searchable.
Duplicates are collapsed. Registries render one skill into .claude/,
.cursor/, .gemini/ and a plugin root, so on a real 60,047-entry catalogue
22,535 rows (37%) are duplicates. Rows whose name and description match, or are
at least 95% alike, fold into one badged ×5, with ^D to show them again.
Identity is the description and never the name: code-reviewer appears 75 times
with 42 different descriptions, and those are 42 real skills.
80 commands, organized into 8 groups
boost --help prints the full grouped list. For every flag of every command see
docs/commands.html; for a visual tour,
docs/index.html; for how boost is put together internally,
the C4 diagrams in docs/architecture/.
| Group | Commands |
|---|---|
| Package Management | install · uninstall · sync · update · reinstall · bundle · import · pin · unpin · snapshot · export · adapt · run |
| Discovery & Search | search · reindex · discover · recommend · browse · index · trending · stats · count |
| Skill Information | list · info · cat · edit · preview · explain · log · home · deps · tag |
| Registry (Taps) | tap · untap · taps · outdated · catalog |
| Intelligence | chat · distill · simulate · infer · absorb · evolve · context · focus · impact |
| Quality & Health | doctor · lint · audit · verify · drift · test · fingerprint · quarantine · decay · heal · conflict · changelog · attest · health · trust |
| Configuration | config · clean · compact · create · policy · onboard · completions · schedule · serve · mcp · hooks · bmad · self-update |
| Team & Collaboration | cohort · profile · protocol · pulse · replay · who |
The AI-assisted commands (search --smart, explain, distill, infer,
absorb, evolve, simulate) use whichever assistant CLI you already have —
claude first, then gemini — or ANTHROPIC_API_KEY, and fall back to plain
heuristics when none is available. boost handles the differences: Gemini takes
no separate system prompt, so it is folded into the message, and a model id
belonging to another vendor is never passed through.
The retrieval and faithfulness evals were measured against Claude, so treat those floors as Claude-measured rather than a claim about every backend.
boost as an MCP server
boost is itself an MCP server, so an agent can search and install skills
mid-task instead of waiting for you to shell out. boost mcp register wires it
into every agent CLI it finds on your PATH, and on a machine with no taps it
pulls in the starter registries first. Installing boost and running boost mcp
is the whole setup.
boost mcp register # every installed host (default: --host auto)
boost mcp register --host gemini # just one
boost mcp register --no-seed # register only; tap nothing
boost mcp unregister # same host selection, in reverse
The seed only fires when the catalog is empty, so re-running boost mcp on a
configured machine leaves your taps alone. Export BOOST_NO_SEED=1 to keep it a
purely local operation, for CI images and air-gapped machines.
| Host | CLI | Registered in |
|---|---|---|
| Claude Code | claude |
claude mcp add → its user-scope MCP config |
| Gemini CLI | gemini |
gemini mcp add -s user → ~/.gemini/settings.json |
Six tools are exposed: boost_search, boost_list, boost_info,
boost_install, boost_doctor and boost_discover_github.
The server also returns MCP instructions, telling the agent when boost is
relevant. The two hosts place those differently, and it matters. Claude Code
loads them into the system prompt. Gemini CLI appends them to its memory tier
alongside GEMINI.md, and only in a trusted folder, so outside one they are
dropped silently. Each tool's own description therefore repeats its trigger,
because a tool description is the only part always in context at the moment the
agent chooses a tool.
Verify with claude mcp list or gemini mcp list; boost should show as
Connected.
Agent hooks and BMAD
boost hooks manages hooks in settings.json for Claude Code and Gemini CLI,
translating between their event vocabularies and timeout units. boost bmad on
uses them to put the BMAD Method's personas in charge of every task, in one
command with no Node and no network.
boost hooks add SessionStart -c 'echo hello' -n greet --scope project
boost bmad on
See docs/bmad.md.
The boost style
The visual identity behind the docs and demos ships as a small dependency-free
design system in style/: the Aurora living-glass look, a
cyan/violet/pink triad on a near-black ground. Drop
style/boost.css into any static page; see
style/README.md for tokens and classes.
The same palette runs in the terminal, in 24-bit color where the terminal
advertises it, degrading to 16-color and then to plain text. The CLI palette is
test-locked to style/boost.css so the two
cannot drift. Color follows the NO_COLOR convention plus
a BOOST_COLOR override: set BOOST_COLOR=never to force plain output, or
always to keep color through a pipe.
Roadmap
Two generated boards track what is shipped, in flight and planned:
code for engine and
tooling work, design
for the Visual Guide and docsite. Both are built from docs/roadmap/items/*.md
and reachable from the Visual Guide.
Working on boost
# everything runs under a disposable HOME, so nothing touches your real setup:
export HOME=/tmp/boost-sandbox && mkdir -p $HOME
python3 tests/make_fixture.py /tmp/fixture-tap
./boost tap /tmp/fixture-tap
./boost install brainstorming
./boost doctor
Every path under ~/.boost and ~/.agents/skills is resolved from $HOME at
call time, which is what makes that sandboxing work.
Four test layers are enforced, and make check runs all of them:
| Layer | Gate |
|---|---|
make test |
≥90% coverage of boost_cli, statements and branches |
make smoke |
176 checks through the actual ./boost shim |
make mutation |
≥80% of ~9,900 mutants in boost_cli/core killed |
make evals |
metric floors on the search ranker, plus no significant regression |
Search quality is the one thing the other three cannot assert: every test can
pass while the ranker quietly gets worse. The eval harness is pure stdlib and
fully offline, generating its own corpus so a metric that moves can only mean
the ranker moved. See evals/README.md.
CONTRIBUTING.md has dev setup, the full gate list, the generated-file rules and the PR process. Start with Good first tasks.
When something misbehaves, boost keeps a rotating log at
~/.boost/logs/boost.log and writes a crash report on any unexpected error.
boost --verbose or boost --debug turn up detail, boost log --diagnostics
reads the trail, and docs/DEBUGGING.md covers the rest.
1 maintainer
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file boost_skill_cli-1.2.40.tar.gz.
File metadata
- Download URL: boost_skill_cli-1.2.40.tar.gz
- Upload date:
- Size: 3.9 MB
- Tags: Source
- Uploaded using Trusted Publishing? Yes
- Uploaded via:
twine/7.0.0 CPython/3.13.14
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
4b959cf69ff0cab84044c1ee02dbfc8d35c7eaa97286d8cee00214603b20b76e
|
|
| MD5 |
a903ea2dd7142960418cbffa6e089c8d
|
|
| BLAKE2b-256 |
48b6f3c8c34aafcf9c13ef9de70d4d0fef16c2ddff23974a6672606fe042d7ee
|
Provenance
The following attestation bundles were made for boost_skill_cli-1.2.40.tar.gz:
Publisher:
publish.yml on jonnyeclectic/boost
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
boost_skill_cli-1.2.40.tar.gz -
Subject digest:
4b959cf69ff0cab84044c1ee02dbfc8d35c7eaa97286d8cee00214603b20b76e - Sigstore transparency entry: 2635478031
- Sigstore integration time:
-
Permalink:
jonnyeclectic/boost@5a18793fa401cf13206f34b4c73abcc9bd8bc323 -
Branch / Tag:
refs/heads/main - Owner: https://github.com/jonnyeclectic
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@5a18793fa401cf13206f34b4c73abcc9bd8bc323 -
Trigger Event:
workflow_run
-
Statement type:
File details
Details for the file boost_skill_cli-1.2.40-py3-none-any.whl.
File metadata
- Download URL: boost_skill_cli-1.2.40-py3-none-any.whl
- Upload date:
- Size: 512.0 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? Yes
- Uploaded via:
twine/7.0.0 CPython/3.13.14
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
e46bcfc24e800fff051c420f465b236908eca386c91dff1cce52f3128f8b2f5a
|
|
| MD5 |
77d3e124c90c0eefcc1565a62b45cc71
|
|
| BLAKE2b-256 |
42837d99c9ecf7fe4c81a73971a5164a286a176812c114369923afb9a1d38a5d
|
Provenance
The following attestation bundles were made for boost_skill_cli-1.2.40-py3-none-any.whl:
Publisher:
publish.yml on jonnyeclectic/boost
-
Statement:
-
Statement type:
https://in-toto.io/Statement/v1 -
Predicate type:
https://docs.pypi.org/attestations/publish/v1 -
Subject name:
boost_skill_cli-1.2.40-py3-none-any.whl -
Subject digest:
e46bcfc24e800fff051c420f465b236908eca386c91dff1cce52f3128f8b2f5a - Sigstore transparency entry: 2635478085
- Sigstore integration time:
-
Permalink:
jonnyeclectic/boost@5a18793fa401cf13206f34b4c73abcc9bd8bc323 -
Branch / Tag:
refs/heads/main - Owner: https://github.com/jonnyeclectic
-
Access:
public
-
Token Issuer:
https://token.actions.githubusercontent.com -
Runner Environment:
github-hosted -
Publication workflow:
publish.yml@5a18793fa401cf13206f34b4c73abcc9bd8bc323 -
Trigger Event:
workflow_run
-
Statement type:
