You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
step 4: drop Context Hub (never used), trim FidgetFlo + add hotlink
- Rename 'Step 4 - FidgetFlo + Context Hub' -> 'Step 4 - FidgetFlo' (+ anchor + Quick Nav + HIW)
- Remove Context Hub subsection + Context Hub rows from install table
- step-4-install.sh: comment out install_context_hub + configure_context_hub_skill calls (functions kept as dead code for reference)
- Collapse Swarm Skills prose into the Commands table (fswarm/fmini/fhive/w4w)
- Fold 'What's an MCP?' and 'Verify MCP' blocks into <details> (keep them but off the main scroll path)
- Hotlink FidgetFlo to lorecraft-io/fidgetflo; 'Want the deep dive?' pointer
Rationale: zero tool_use invocations of get-api-docs / chub / contextHub in any session
log under ~/.claude/projects/; MCP not registered in ~/.claude.json; skill file stale
(unchanged since Mar 23). Keeping it in docs was misleading.
@@ -88,7 +88,7 @@ Run the steps in order. Each one builds on the last.
88
88
89
89
**[Step 3 — Developer & Utility Tools](#step-3---developer--utility-tools)** is where you install the rest of your development tools. Things like file converters, search tools, and utilities. You run this from your terminal after Step 1 is done. Much more straightforward.
90
90
91
-
**[Step 4 — FidgetFlo + Context Hub](#step-4---fidgetflo--context-hub)** is where you set up FidgetFlo and Context Hub. FidgetFlo is the multi-agent orchestration layer that turns Claude into a full team of AI agents. Context Hub stops Claude from hallucinating when writing code that calls APIs.
91
+
**[Step 4 — FidgetFlo](#step-4---fidgetflo)** is where you set up [FidgetFlo](https://github.com/lorecraft-io/fidgetflo), the multi-agent orchestration layer that turns Claude into a full team of AI agents — `/fswarm`, `/fmini`, `/fhive`, persistent memory, Opus-locked.
92
92
93
93
**[Step 5 — Productivity Tools](#step-5---productivity-tools)** connects Claude to your productivity tools — notes, calendars, meetings, workflows, browser automation, and hosted toolkits. Pick the ones you use: Notion, Granola, your own n8n instance, Google Calendar, Morgen (recommended), Motion Calendar, Playwright, or SwiftKit. All optional, install only what you need.
94
94
@@ -346,126 +346,87 @@ A "stop hook" fires every time you end a Claude session (Ctrl+C or `/exit`). Cla
346
346
347
347
---
348
348
349
-
## Step 4 - FidgetFlo + Context Hub
349
+
## Step 4 - FidgetFlo
350
350
351
351
[Back to top](#quick-navigation)
352
352
353
-
This step installs [**FidgetFlo**](https://github.com/lorecraft-io/fidgetflo), a multi-agent swarming layer that turns Claude from a single assistant into a coordinated team of AI agents, plus [Context Hub](https://github.com/andrewyng/context-hub), which stops those agents from hallucinating when writing code that calls external APIs.
354
-
355
-
### FidgetFlo
356
-
357
-
[**FidgetFlo**](https://github.com/lorecraft-io/fidgetflo) is a fork of [ruvnet's Ruflo](https://github.com/ruvnet/ruflo), tuned for Claude Opus 4.7. It adds to Claude Code:
- **Opus locked by default** — no silent downgrades to Haiku or Sonnet. Every spawned agent runs on Opus.
361
-
- **Persistent agent memory** — context survives across sessions, agents share what they learn.
362
-
- **Self-healing workflows** — failed steps detect + recover instead of just stopping.
363
-
364
-
Want the deep dive? Architecture, agent catalog (60+ types), memory system, hook pipeline, topology options — all in the [FidgetFlo repo →](https://github.com/lorecraft-io/fidgetflo)
365
-
366
-
### Context Hub
367
-
368
-
Built by [@andrewyng](https://github.com/andrewyng) ([repo](https://github.com/andrewyng/context-hub)). Andrew Ng is one of the most respected names in AI. He built this tool to solve a real problem: AI agents hallucinating API calls.
369
-
370
-
When Claude writes code that calls an external API, it's working from its training data, which can be outdated or just wrong. Context Hub fixes that by giving Claude access to curated, up-to-date API documentation on demand.
371
-
372
-
- **Accurate API docs.** Claude can look up the real function signatures, parameters, and usage patterns instead of guessing from memory.
373
-
- **Persistent annotations.** You and Claude can add notes to docs that carry over across sessions. If you figure out a quirk with an API, it stays documented.
374
-
- **Less hallucination.** This is the big one. Claude stops making up function names that don't exist.
375
-
376
-
Together, FidgetFlo and Context Hub are what take you from "using AI" to actually being an AI super user.
377
-
378
-
### Swarm Skills
379
-
380
-
Step 4 also installs three slash commands that let you launch multi-agent swarms on demand:
381
-
382
-
- **`/fswarm <task>`** — Launches 15 agents immediately. You describe the task, Claude assigns roles (architect, coders, testers, security auditor, etc.) and they all work in parallel. Use this when you know what you want done and want brute-force execution.
383
-
384
-
- **`/fmini <task>`** — Launches 5 agents immediately. Same architecture as `/fswarm` but with a tighter team: architect, developer, tester, reviewer, and researcher. Use this for focused tasks where you don't need the full battalion.
385
-
386
-
- **`/fhive <goal>`** — Launches a queen agent that autonomously manages everything. You describe the goal, the queen decides how to break it down, what workers to spawn, and how to coordinate them. Use this when you want to set a direction and step back.
387
-
388
-
#### Swarm tiers — extended thinking
389
-
390
-
Subagents inherit the parent session's model (Opus stays Opus) but do **not** inherit its extended-thinking setting. The `/effort` slider does not tether to spawned agents, so the only way to get deep thinking inside a swarm is to bake a trigger phrase into each Agent's prompt. That's what these tier variants do:
Base `/fswarm` and `/fmini` are unchanged — the tier variants are purely additive. Use `/fminimax` for architecture reasoning or hairy debugging where you want a small team thinking hard; use `/fmini2` for routine feature work that still benefits from a bit of deliberation; reach for `/fswarm3` or `/fswarmmax` when you need 15 agents each chewing on a problem at full depth.
403
-
404
-
All 10 commands enforce Opus-only (never Haiku or Sonnet) and all share the same FidgetFlo `/tmp/` signal files, so the statusline indicators (🐝 for `/fswarm*`, 🍯 for `/fmini*`) fire for every tier.
405
-
406
-
All three skills signal the statusline so you can see live indicators while agents are working.
353
+
[**FidgetFlo**](https://github.com/lorecraft-io/fidgetflo) is a fork of [ruvnet's Ruflo](https://github.com/ruvnet/ruflo), tuned for Claude Opus 4.7. It turns Claude Code from a single assistant into a coordinated team of AI agents: multi-agent swarms on demand, persistent memory, self-healing workflows, and all agents Opus-locked by default (no silent downgrade to Haiku/Sonnet).
407
354
408
355
### Run Step 4
409
356
410
-
You should still have a Claude session open from Step 3. If you closed it, open your terminal and type `cskip` to start a new Claude session. Remember, you can press **Shift+Tab** at any time to toggle auto-approve on or off.
411
-
412
-
Once you're inside the Claude session, paste this and hit Enter:
357
+
Still in a `cskip` session? Good. Paste this:
413
358
414
359
> [!IMPORTANT]
415
360
> **Paste this into your Claude session:**
416
361
> ```
417
-
> run this command to set up FidgetFlo and Context Hub: bash <(curl -fsSL https://raw.githubusercontent.com/lorecraft-io/cli-maxxing/main/step-4/step-4-install.sh)
362
+
> run this command to set up FidgetFlo: bash <(curl -fsSL https://raw.githubusercontent.com/lorecraft-io/cli-maxxing/main/step-4/step-4-install.sh)
418
363
> ```
419
364
365
+
If Claude tells you to restart your terminal, close the window, reopen, `cskip` again, and tell Claude to pick up where it left off.
366
+
367
+
### Commands
368
+
369
+
| Command | What it does |
370
+
|---------|-------------|
371
+
| `/fswarm <task>` | Launches 15 agents in parallel — architect, backend devs, testers, security auditor, etc. Brute-force execution. |
372
+
| `/fmini <task>` | Launches 5 agents — architect, dev, tester, reviewer, researcher. Tighter team for focused work. |
373
+
| `/fhive <goal>` | Queen agent takes full control — decides what workers to spawn and how to coordinate. Set the goal, step back. |
374
+
| `/w4w` | Word-for-word, line-for-line. Maximum attention, zero skipping. |
420
375
421
-
Claude will run the install for you. Same as Step 3. If Claude tells you to restart your terminal, close the window, reopen your terminal, type `cskip`, and let Claude know where you left off.
376
+
#### Thinking tiers
422
377
423
-
### Quick note: what's an MCP?
378
+
Subagents inherit the parent's model (Opus stays Opus) but **not** its `/effort` setting — the slider doesn't tether to spawned agents. These tier variants bake the trigger into each Agent prompt:
424
379
425
-
You'll see "MCP" mentioned here and in future steps. MCP stands for Model Context Protocol. Think of it as a plugin system for Claude. When you connect an MCP to Claude Code, you're giving Claude access to a new tool or data source it didn't have before. Claude can then use that tool automatically whenever it's relevant. You don't have to manage it. You just connect it once and Claude takes it from there.
Natural-language aliases work too: "hard"/"deep" → tier 2, "harder"/"deeper" → tier 3, "max"/"mega"/"ultra" → max. All variants are Opus-only and fire the 🐝 `/fswarm*` / 🍯 `/fmini*` statusline indicators.
426
389
427
390
### What This Step Installs
428
391
429
392
| Component | What it does |
430
393
|-----------|-------------|
431
-
| FidgetFlo CLI | This is the command-line tool that manages everything below. |
432
-
| MCP Server Connection | This connects FidgetFlo to Claude Code so they can talk to each other (see above). |
433
-
| FidgetFlo Daemon | This runs in the background and coordinates agents, memory, and tasks. |
434
-
| Memory System | This gives your agents persistent, searchable memory across sessions. |
435
-
| Smart Model Routing | This is disabled by default so you always get Opus. Can be turned on later to save up to 75% on costs by routing simple tasks to cheaper models. |
436
-
| Opus Lock | This locks all tasks to Opus so nothing silently downgrades to a weaker model. You're paying for Opus, so you should always get Opus. |
437
-
| Context Hub | This gives Claude access to curated, up-to-date API documentation so it stops hallucinating function names. |
438
-
| Context Hub Skill | This teaches Claude when and how to look up API docs automatically before writing integration code. |
439
-
| Swarm Skill (`/fswarm`) | Type `/fswarm` followed by any task and Claude immediately launches 15 parallel agents to tackle it. A lead architect coordinates backend devs, testers, security auditors, and more — all working simultaneously. |
440
-
| Mini Swarm Skill (`/fmini`) | Type `/fmini` followed by any task and Claude launches 5 focused agents — architect, developer, tester, reviewer, and researcher. Same parallel execution as `/fswarm`, just a tighter team for focused work. |
441
-
| Hive Skill (`/fhive`) | Type `/fhive` followed by a goal and a queen agent takes full control. She decides what workers to spawn, how to coordinate them, and when the work is done. You set the goal and step back. |
442
-
| Attention Skill (`/w4w`) | Word-for-word, line-for-line mode. Maximum attention to detail for critical tasks. |
443
-
| TypeScript | Required by some FidgetFlo features. Installed globally. |
444
-
| agentic-flow | Enables embeddings and advanced routing capabilities. |
445
-
| Statusline | A real-time status bar that shows your active tools, model, session time, and context usage. When a swarm or hive is running, you'll see live indicators so you always know what's happening. |
394
+
| FidgetFlo CLI + daemon | Coordinates agents, memory, and tasks in the background. |
395
+
| MCP Server | Wires FidgetFlo into Claude Code. |
396
+
| Memory System | Persistent, searchable memory shared across agents + sessions. |
397
+
| Opus Lock | All tasks and spawned agents run on Opus — no silent downgrade to Haiku/Sonnet. |
| Statusline | Live indicators for swarms, hives, model, session time, and context usage. |
401
+
402
+
**Want the deep dive?** Architecture, agent catalog (60+ types), memory system, hook pipeline, topology options — all in the [FidgetFlo repo →](https://github.com/lorecraft-io/fidgetflo)
446
403
447
404
### After Step 4
448
405
449
-
Your core tools are installed. Continue to Step 5 for productivity tools. Or open a new `cskip` session and try something ambitious. FidgetFlo kicks in automatically when the task calls for it.
406
+
Your core tools are installed. Continue to Step 5 for productivity tools — or open a new `cskip` session and try something ambitious. FidgetFlo kicks in automatically when the task calls for it.
450
407
451
-
### MCP Server Setup
408
+
<details>
409
+
<summary><strong>What's an MCP?</strong></summary>
452
410
453
-
Claude Code can connect to MCP (Model Context Protocol) servers for extended capabilities. After running Step 4 (FidgetFlo), the MCP server is configured automatically.
411
+
MCP stands for Model Context Protocol — a plugin system for Claude. Connect an MCP once and Claude gets access to a new tool or data source, which it picks up automatically when relevant.
454
412
455
-
For manual MCP setup or troubleshooting, see the [Claude Code MCP documentation](https://docs.anthropic.com/en/docs/claude-code/mcp-servers).
413
+
</details>
456
414
457
-
#### Verify MCP Connection
415
+
<details>
416
+
<summary><strong>Verify the FidgetFlo MCP is connected</strong></summary>
458
417
459
-
After setup, verify the MCP server is connected:
460
418
```bash
461
419
claude mcp list
462
420
```
463
421
464
-
If the FidgetFlo MCP server isn't showing, re-add it:
422
+
If FidgetFlo isn't listed, re-add it:
423
+
465
424
```bash
466
425
claude mcp add fidgetflo -- npx -y fidgetflo@latest
0 commit comments