v0.6.2 — Instagram MCP server bundled inside package

Instagram MCP server bundled inside ai-bootstrap package — works on any computer.

Fixed (the real fix)

v0.6.1 added the integration command but the Instagram MCP server source was
NOT inside the ai-bootstrap package. Users on a fresh machine had nothing to install.

v0.6.2 bundles the full 15-tool Instagram MCP server (5 source files + package.json

• .env.example + README + meta-app-config docs) inside the npm tarball.

Architecture

• New: packages/templates/integrations/<name>/ — source-of-truth for local MCP servers
• prepack (copy-templates.mjs) now includes integrations/ alongside skills/agents
• ai-bootstrap integration install <name>:
  1. Looks up bundled source from package's templates/integrations/<name>/
  2. Copies to ~/.claude/integrations/<name>/
  3. Runs npm install in server/
  4. Seeds .env from .env.example (idempotent — won't overwrite existing)
  5. Prints next-steps (fill .env, doctor, mcp add)

Workflow on a fresh machine

npm install -g @azerogluemin/ai-bootstrap@latest
ai-bootstrap                                    # foundation + free MCPs
ai-bootstrap integration install instagram      # copy + npm install + seed .env
nano ~/.claude/integrations/instagram-mcp/.env  # add Meta App ID + token
ai-bootstrap integration doctor instagram       # 5/5 health check
ai-bootstrap mcp add instagram                  # bind to Claude Code
claude                                          # ig_profile etc. available

Tarball impact

• v0.6.1 → v0.6.2: +9 files, ~26 KB (server code + docs)
• Still under 1.3 MB total

v0.6.1 — Instagram MCP fix (local server) + integration command

Real Instagram MCP fix + new integration command for local MCP servers.

Fixed

• Instagram MCP catalog entry was fabricated: pointed to @modelcontextprotocol/server-instagram which does NOT exist on npm (Rule 16 violation discovered in real testing).
• Now uses a LOCAL Node server at ~/.claude/integrations/instagram-mcp/ (15-tool real Graph API implementation: ig_profile, ig_publish_reel, ig_list_messages, etc.). Server is shipped separately and copied from a working source (e.g., azerogluemin.az/projects/azerogluemin-ai/instagram-mcp).
• Env vars (META_APP_ID, IG_ACCESS_TOKEN, IG_BUSINESS_ACCOUNT_ID, etc.) are loaded by the server itself from ~/.claude/integrations/instagram-mcp/.env — no shell env required.

Added — ai-bootstrap integration <sub> command

For MCP servers that ship as local Node code instead of public npm packages:

• ai-bootstrap integration status — list installed integrations + health
• ai-bootstrap integration install <name> — npm install deps + verify .env + first boot
• ai-bootstrap integration doctor <name> — full health check (file/deps/env/boot)

Currently supports: instagram (15-tool local server). Architecture is extensible — drop a new server into ~/.claude/integrations/<name>/ and register it in INTEGRATIONS map.

Why

Some valuable MCP integrations exist only as project-local code (custom Graph API clients, internal tools). ai-bootstrap now first-classes that pattern instead of forcing everything through the npm registry.

v0.6.0 — preset-based bootstrap + 8 new skills + folder scaffolding

Preset-based project bootstrap — folder scaffolding + CLAUDE.md per preset.

Added — 8 new production skills

Real production-quality content (~150-280 lines each, with sources):

Dev specialists:

• frontend-developer — React 19/Next 16/Vue 3/Svelte 5, state mgmt, performance, WCAG 2.2, testing
• backend-developer — REST/GraphQL, NestJS/Hono/FastAPI, multi-tenant, OWASP, caching, queues
• devops-developer — Docker, K8s, CI/CD, IaC, secrets, monitoring, SLOs
• mobile-developer — React Native (New Arch), Flutter, native iOS/Android, offline sync

Film production specialists:

• editor — Cut grammar, J/L cuts, vertical pacing, sound editing, Resolve/Premiere/FCP/CapCut
• colorist — Scopes (waveform/vector/parade), LUT design, Pixar/Wong Kar-wai/Deakins/A24 looks
• art-director — Visual treatment, mood board, palette per scene/character, era research (AZ-specific)
• cinematographer — Camera/lens/light decision trees, Deakins/Lubezki/Fraser craft

Added — 3 project presets

Replaces bundle abstraction in user-facing CLI. Each preset includes
folder structure + CLAUDE.md template + skill+agent install list + MCP suggestions.

SaaS Development (~58 skills, ~50 agents, ~22 MCPs)

• Folders: `apps/{web,api,admin}`, `packages/{ui,db,shared}`, `docs/00-27`, `infra/`, `scripts/`
• CLAUDE.md rules: TS strict, multi-tenant, JWT+refresh, OWASP, decisions-log append-only
• 28 numbered docs scaffolded (00-claude-code-guide → 27-go-live-runbook)

Social Page (~44 skills, ~22 agents, ~21 MCPs)

• Folders: `strategy/`, `calendar/`, `days/YYYY-MM-DD/{characters,shots,video,music,thumbnail,final}`,
  `prompts-library/`, `assets/`, `analytics/`
• CLAUDE.md rules: Prompt-first (vizual təsdiqsiz fayl qadağa), daily folder isolation
• Strategy MD skeletons: brand-guide, content-pillars, audience-personas, competitors

AI Studio (~34 skills, ~11 agents, ~14 MCPs)

• Folders: `projects//{brief,characters,locations,shots,video,music}`,
  `days/YYYY-MM-DD/`, `references/{pixar-style,cinematography,color-grading}`,
  `prompts-library/`
• CLAUDE.md rules: Prompt-first, style refs before prompt, file naming convention
• NO social platform skills/MCPs (studio doesn't publish)

Changed — ai-bootstrap new is now preset-based

Previous (v0.5.0): multi-select bundle checkbox
New (v0.6.0): single-select preset → 3 questions only:

1. Layihə adı?
2. Bu qovluqda nə qurmaq istəyirsən? (SaaS / Social Page / AI Studio)
3. Qısa təsvir?

After selection: skills+agents symlinked from pool, MCPs auto-suggested,
folder structure scaffolded, CLAUDE.md written with preset-specific rules.

Architecture

• New: `packages/cli/src/applier/preset-definitions.ts` — 3 preset metadata
• New: `packages/cli/src/applier/preset-scaffolder.ts` — folder scaffolder per preset
• Bundles still exist internally (`resolvePlan`, `SKILL_BUNDLES`, `AGENT_BUNDLES`)
  but no longer surface to users — presets replaced them
• Project state schema bumped to v2.0: `{ preset, name, description, createdAt }`

Tests

• 176 passing (was 142): 121 smoke + 55 e2e
• New: 30+ preset definition assertions (skill counts, MCP presence, preset uniqueness)
• New: 11 scaffolder assertions (SaaS 28 docs, Social strategy MDs, AI Studio refs)

Migration

• v0.5.0 projects (multi-bundle): continue to work; new `ai-bootstrap new` writes preset state
• Old multi-bundle CLAUDE.md template still readable; project state v2.0 backward compat

v0.4.1 — README rewrite published to npm

README rewrite — published so npm.com lands on the new usage guide.

Changed

• packages/cli/README.md — full rewrite (AZ):
  • 2-step TL;DR quick start
  • User scope vs. project scope explained with table
  • 5 real workflow scenarios (Creator/IG, SaaS, SMM, add bundle, remove)
  • All commands grouped by purpose (setup, mgmt, MCP, backup, telemetry)
  • 63 MCP server categorized list
  • Troubleshooting section (5 common issues)
• No functional code changes vs. 0.4.0.

v0.4.0 — add/remove/list: incremental skill management

Incremental skill + agent management — add / remove / list.

Added

• ai-bootstrap add <names> — install additional skills or agents
  • Auto-detects skill vs agent by template lookup
  • --bundle <name> merges an entire bundle
  • Auto-detects scope (project if CLAUDE.md / .claude/ in cwd, else user)
  • --user / --project flags force scope
• ai-bootstrap remove <names> (alias: rm) — uninstall skills/agents
  • Same scope auto-detection
  • Removes from both skills/ and agents/ if name matches both
• ai-bootstrap list (alias: ls) — list installed skills/agents
  • Current scope by default
  • --user, --project, or --all flags
  • --all shows user scope + current project side-by-side

Why

v0.3.0 introduced new for project bootstrap with a single bundle. But real
users hit edge cases:

• Picked developer bundle, later need showrunner for a marketing page
• Picked creator, later need seo-technical for SEO audit
• Want to remove unused skills to reduce context noise

add / remove / list give granular control without re-running the whole wizard.

Workflow examples

# Already have a developer-bundle project, need creator skills too:
ai-bootstrap add showrunner character-designer image-prompt-engineer

# Or merge whole creator bundle on top:
ai-bootstrap add --bundle creator

# See what's installed in this project:
ai-bootstrap list

# See user + project side by side:
ai-bootstrap list --all

# Drop a skill you don't use:
ai-bootstrap remove security-auditor

Tests

• New e2e tests: add help, list empty user scope, remove non-existent graceful
• 134 passing (was 131): 87 smoke + 47 e2e

v0.3.0 — ai-bootstrap new: project-scope skills

Project-scope skills + agents — different bundles per project.

Added

• ai-bootstrap new subcommand — bootstrap any folder as a project:
  1. Ask the project name (defaults to folder name)
  2. Ask "what is this project?" with 9 intent choices that auto-map to a bundle:
     • SaaS / Fullstack → developer
     • AI Creator content → creator
     • Marketing / SMM → marketer
     • Mobile app → developer
     • Data analysis / dashboard → developer
     • Client agency work → full-stack
     • Startup / founder → founder
     • Open source library → developer
     • Just basics → foundation
  3. Allow bundle override
  4. Ask for 1-2 sentence project description (goes into CLAUDE.md)
  5. Allow custom project-specific rules
  6. Install skills to <project>/.claude/skills/ (PROJECT scope)
  7. Install agents to <project>/.claude/agents/ (PROJECT scope)
  8. Write <project>/CLAUDE.md with project metadata + bundle reference
  9. Save state to <project>/.claude/ai-bootstrap-project.json

Why this matters

Previously, ai-bootstrap installed ONE bundle globally (~/.claude/). Every
project saw the same skill set. Real users have multiple identities — AI
creator on Monday, full-stack dev on Tuesday, marketer on Wednesday — and
need different skills per project.

Claude Code's project-scope skills (<project>/.claude/skills/) load ONLY
when working inside that project. This gives you the right tools per context
without polluting every session.

Changed

• installSkills(names, targetDir?) — second arg defaults to ~/.claude/skills/
  for backward compat; project install passes <project>/.claude/skills/
• installAgents(names, targetDir?) — same pattern
• Help text reorganized: user-scope wizard vs project-scope new

Testing

• New smoke test: 'project-scope install' — verifies skills install to
  project dir, user-scope dir NOT affected
• E2E test updated for help text changes
• 127 tests passing (was 124)

Workflow

# Once per machine — universal foundation:
npx @azerogluemin/ai-bootstrap   # picks foundation/minimal user-scope

# Then per project:
cd ~/Projects/my-saas && ai-bootstrap new     # → developer bundle
cd ~/Projects/azerogluemin-content && ai-bootstrap new   # → creator bundle
cd ~/Clients/restaurant-smm && ai-bootstrap new   # → marketer bundle

Each project sees only its own bundle. No skill pollution.

v0.2.2 — wizard URL + public repo user-autonomy

User-autonomy fixes from first real-world test feedback.

Fixed

• Wizard step 6 (GitHub backup) over-validated URLs:
  • Now accepts both SSH (git@github.com:user/repo) and HTTPS
  • Empty input is now allowed → defer URL to ai-bootstrap backup init later
  • User no longer locked into giving a URL at wizard time
• backup init refused public repos outright:
  • Now warns + asks for explicit confirmation instead of hard-blocking
  • Rationale: user autonomy — the user knows their threat model

Why

In v0.2.1 test, the wizard rejected a valid SSH URL and refused to proceed.
Two design errors: (1) regex was https-only, (2) URL was mandatory if user
said "yes" to backup. Both fixed.

v0.2.1 — first npm publish

First published version on npm.

Changed

• No functional code changes vs. 0.2.0.
• Released manually via npm publish from a logged-in local environment
  (the v0.2.0 commit existed only on GitHub; this is the npm-first release).

v0.2.0 — MCP catalog 63, backup + telemetry, 124 tests

0.2.0 — 2026-06-20

Second release — closes the "post-launch roadmap" gaps from v0.1.0.

Added

• MCP catalog expanded 41 → 63 servers (all verified against official docs or
  npm/GitHub as of 2026-06-20; no fabricated packages):
  • Atlassian (Jira + Confluence) via uvx mcp-atlassian (sooperset)
  • Twilio via @twilio-alpha/mcp (official alpha)
  • SendGrid via sendgrid-mcp (Garoth)
  • Trello via @delorenj/mcp-server-trello
  • Asana via mcp-server-asana
  • ClickUp via clickup-mcp-server
  • Spotify via spotify-mcp-server (marcelmarais)
  • YouTube Transcript via @sinco-lab/mcp-youtube-transcript
  • arXiv via uvx mcp-simple-arxiv
  • Kubernetes via docker ghcr.io/alexei-led/k8s-mcp-server
  • Browserbase via @browserbasehq/mcp
  • Apify via @apify/actors-mcp-server (4500+ scrapers)
  • Time via uvx mcp-server-time (no creds, timezone utils)
  • Google Workspace official remote MCPs (HTTP transport + OAuth):
    Gmail, Google Drive, Google Calendar, Google Chat
• HTTP transport support in catalog schema:
  • transport: 'stdio' | 'http', serverUrl, oauth: { clientIdEnv, clientSecretEnv }
  • writeMcpConfig writes either stdio command/args/env OR serverUrl/oauth JSON
• New CLI command: ai-bootstrap backup <sub>
  • backup init — git init + remote + initial push (uses system git auth)
  • backup sync — stage + commit + push current state
  • backup pull — clone or pull from remote (restore on new machine)
  • backup status — show remote, last sync, last commit
  • Auto-generates .gitignore excluding mcp-tracking.json (no creds leaked)
• New CLI command: ai-bootstrap telemetry <sub>
  • telemetry on/off/status — strict opt-in (default off)
  • Writes ~/.claude/telemetry.json with random install UUID
  • Endpoint configured via AI_BOOTSTRAP_TELEMETRY_URL env (no default —
    no data goes anywhere unless user sets a collector)
  • Collects: version, Node, platform, bundle choice, MCP IDs, event name
  • Never collects: profile, project names/paths, knowledge files, credentials
  • 2-second timeout, silent failure (never breaks CLI)
• E2E test suite (tests/e2e.test.mjs): 43 tests covering compiled CLI
  via subprocess — --version, --help, doctor on empty HOME, mcp list/installed,
  update with no state, telemetry status/on/off, backup status, error handling
• New category: research for arxiv, youtube-transcript

Changed

• package.json scripts:
  • test = smoke + e2e (was: smoke only)
  • test:smoke and test:e2e for individual runs
  • prepublishOnly now runs e2e too before publish
• Help text reorganized to show new subcommands

Quality

• 124 tests passing (was 81): 81 smoke + 43 e2e
• All MCP catalog entries include source: field attributing the maintainer
  (official vs community) for transparency
• HTTP transport entries clearly marked with transport: 'http'

Known limitations (still post-launch)

• Documentation site (ai-bootstrap.dev) — separate marketing project
• Some MCPs depend on npm packages that may rename; catalog will need refresh
• Telemetry collector backend is user-provided (intentional — privacy-first)
• ProductHunt / HackerNews launch — manual