SECTION · BUILT WITH · 16 / 20

What we built it with.

Honest receipts for the ten off-the-shelf parts the crew runs on — no partnership, no sponsorship, no hand-waving. The substrate on top of them is what makes RHOBEAR different (see substrate.html); this page is just the parts list.

Every card opens with what you get from it inside RHOBEAR. Expand any one to see the spec, the receipts, and how it got here. Logos are the real marks. Star counts are public GitHub snapshots. No partnership, sponsorship, or endorsement implied.
10 ingredients6 MIT · 3 Apache-2.0 · 1 closed-source-paired~803,000 combined GitHub starsAll swappable in .env
↓ The room these ten parts run inside. ↓
The RHOBEAR hub — Starship 17-B
Claude Code "The plan-then-execute split that maps onto Architect and Foreman without a fight." ★ 129,446 Apache-2.0 github.com/anthropics/claude-code DETAILS How it lives inside RHOBEAR Give Pops a sentence — Pops opens a Claude Code session. Send Architect a multi-file refactor; plan mode drafts the lanes before a byte changes. Foreman runs Claude Code execute mode behind the autoloop while you sleep.

The power

  • Reads a real shipping codebase without flattening it
  • Multi-file edits with a diff you can argue with before it lands
  • Plan mode drafts the work, execute mode does it — same session
  • Tool-use that waits for the file to be there before pretending to edit

Why we chose it

  • Does the unglamorous part well: find the file, edit surgically, explain the diff
  • Session log feeds straight into worker-sessions
  • Plan/execute split maps one-to-one onto Architect and Foreman

Evolution

  • 2024 Q3 — released as CLI companion to Sonnet 3.5
  • 2025 — Sonnet 4 brought clean multi-file edits
  • 2026 — Sonnet 4.5 makes plan-mode the default, autoloop stabilizes

Where you might know it from

  • The CLI you launch with claude after npm install
  • The "plan mode / execute mode" agent everyone has tried
  • Anthropic's first-party answer to Cursor and Codex
Codex openai/codex "A second engine to put head-to-head against Claude — pick the winner per task." ★ 87,789 Apache-2.0 github.com/openai/codex DETAILS How it lives inside RHOBEAR Set FOREMAN_PROVIDER=codex — same Foreman role, different model lineage. Run Codex on one lane, Claude on another in head-to-head — promote the winning diff. Lean on Codex for heavy refactor planning and schema work.

The power

  • Apache-2.0 Rust — fast cold start, no Node dependency tree
  • Native tool-use: shell, file, search, structured patches
  • GPT-5 Codex reasoning with terminal-native discipline
  • Pushed near-daily — actively maintained inside OpenAI

Why we chose it

  • Provider-agnostic Foreman needs a strong second rail next to Claude Code
  • Apache-2.0 lets us ship inside the Tauri installer without a license fight
  • Different model lineage from Claude — head-to-head surfaces real differences

Evolution

  • 2021 — original Codex model behind the original GitHub Copilot
  • 2025 — repo relaunched as first-party Rust CLI under openai/codex
  • 2026 — GPT-5 Codex becomes default model, tool-use stabilizes

Where you might know it from

  • OpenAI's official Rust CLI for GPT-5 Codex
  • The terminal agent that ships from OpenAI org, not a third-party wrapper
  • The "I just want a focused terminal coder" answer from OpenAI
↓ Three engines in three panes. Same workbench. ↓
RHOBEAR CLI wall — Claude Code, Pi, shell, log tail
~$14.70
saved per 100-turn session
50×
smaller system prompt
Flat-cost subscription · same Codex model family · less per-turn overhead
Pi "A flat-cost coding lane next to your metered ones — cheap builder, expensive planner." npm · @earendil-works/pi-coding-agent MIT npmjs.com/.../pi-coding-agent DETAILS How it lives inside RHOBEAR Set FOREMAN_PROVIDER=pi — Foreman binds to a Pi session billed against your existing ChatGPT Plus sub. Run bulk agentic work (CSV transforms, codemods, mass renames) on Pi — the per-token bill stops climbing. Pair Pi (Foreman) with Claude (Architect) — the cheap builder ships, the expensive planner thinks.

The power

  • Subscription-billed instead of per-token — bill stops climbing
  • ~200-token system prompt vs Claude Code's ~10,000 — less overhead per turn
  • Same Codex/GPT-5 model family Claude users already trust
  • MIT — auditable, forkable, no rug-pull risk

Why we chose it

  • Lets RHOBEAR offer a flat-cost lane next to metered ones
  • Lighter system prompt saves real dollars at the same task
  • Shares core with OpenClaw (Earendil-works lineage), clean integration

Evolution

  • 2025 — initial release as a thin terminal coder targeting OSS Codex
  • 2026 H1 — OpenAI publicly endorsed it as a recommended OSS path
  • 2026 — Spark/Codex-quota integration lets it ride a Plus/Pro sub

Where you might know it from

  • Earendil-works coding agent that ships with the Codex-OAuth ChatGPT Plus already pays for
  • The "flat-cost coder" people switched to when API bills started showing up
  • OpenAI-endorsed sibling project for OSS Codex tooling

npm-published, not a starred GitHub project — flat star count would be misleading. The package weekly-downloads number on npm is the right proxy if you want one.

Whisper "Talk to your hub — and the audio never leaves your machine." ★ 101,199 MIT github.com/openai/whisper DETAILS How it lives inside RHOBEAR Click the mic on the chat dock — talk for up to 30 seconds — transcript drops into your active input. Dictate a goal into Compass while you walk the dog — the autoloop picks it up when you sit down. Spark uses it for audio-job captioning so every render ships transcribed.

The power

  • On-device — audio never leaves the machine
  • ggml-base.bin is ~145MB, runs warm in 300-800ms on a normal laptop
  • MIT — fork it, bundle it, ship it
  • Multilingual without a separate model per language

Why we chose it

  • The "talk to your hub" workflow only works if dictation is local
  • Downloads on first-run consent — installer stays at ~58MB
  • MIT removes every licensing question before it gets asked

Evolution

  • 2022 Sep — initial OpenAI release as a research-grade speech model
  • 2023 — ggml port (whisper.cpp) for consumer hardware
  • 2024-2025 — base/small/medium variants stabilize as default local stack

Where you might know it from

  • The 2022 OpenAI speech model that made on-device transcription real
  • The ggml-base.bin file in every "talk to your computer" tutorial
  • The reason your Mac's local dictation got tolerable in 2023
Ollama "Your AI, fully local — nothing leaves the box when you say so." ★ 172,897 MIT github.com/ollama/ollama DETAILS How it lives inside RHOBEAR Set every role's provider to ollama in .env — the crew binds to local models. Run Pi on a paid lane, Ollama on a free lane, compare in head-to-head. Boot on a $5 VPS with Ollama and run fully API-free.

The power

  • MIT — fully open, fully local
  • Single-binary runtime for Llama 3.2, Qwen, Mistral, and ~100 other open weights
  • Net-egress can be turned off and the runtime keeps working
  • Same CLI shape as docker run — easy mental model

Why we chose it

  • "Your AI" promise has a floor: nothing leaves the box when you say so
  • Every role in the hub contract has an ollama provider option
  • Safety registry can pin net:egress=off and the autoloop runs slower instead of failing

Evolution

  • 2023 — initial release as a Mac-first local runner
  • 2024 — Windows + Linux first-class support, library explodes
  • 2025-2026 — Llama 3.2 and Qwen 2.5 ship same-day as Ollama recipes

Where you might know it from

  • The ollama run llama3.2 command that made local models a one-liner
  • The runtime that took the friction out of "I just want to try a model on my laptop"
  • The thing that turned "run a model offline" from a weekend project into a CLI command
MCP Model Context Protocol "The USB-C of AI tools. Plug a tool in once — the whole crew can use it." ★ 86,599 (servers) · 8,295 (spec) MIT github.com/modelcontextprotocol DETAILS How it lives inside RHOBEAR Wizard installs an MCP server from the catalog — every authorized role can call it immediately. Drop a Stripe MCP in — Foreman can issue refunds, Guardian can audit them. Build your own MCP server for an internal API — the crew picks it up with no adapter code. Business-vertical skills (Legal, Marketing, Finance, HR, Sales, Brand Voice) live in the same catalog — see the Skills strip on the home page.

The power

  • Write a tool server once, every MCP-aware agent in the crew can call it — no per-model rewrites, ever.
  • Kills the N×M integration nightmare: one tool, every model, instead of re-wiring fs/shell/search for each provider separately.
  • The standard the whole industry converged on in ~14 months — Anthropic, OpenAI, Google, Microsoft, AWS. You're building on the agreed-upon plug, not a proprietary bet that dies in a year.
  • Per-role scope (fs:read vs fs:write vs shell:exec) enforced by the protocol — structured, auditable, serializable.

Why we chose it

  • Re-implementing tools per model provider was a non-starter — MCP means we never do it twice.
  • "Install a tool, every agent can use it" maps one-to-one onto Wizard's job.
  • Per-role scope is exactly how the safety registry wanted to work anyway.

Evolution

  • 2024 Nov — Anthropic publishes the original spec and reference servers
  • 2025 — OpenAI, Google, and major IDE vendors adopt it
  • 2026 — community catalog crosses 1,000 servers

Where you might know it from

  • Anthropic's tool-bridge standard adopted by every other agent product in 2024-2025
  • The reason fs/shell/search tools work across multiple model providers without rewrites
  • The mcp-server-* packages flooding npm and PyPI
Aguara "A reading-distance security scanner — small enough that we read every rule before shipping it." ★ 81 Apache-2.0 github.com/garagon/aguara DETAILS How it lives inside RHOBEAR Every worker diff runs through Aguara in the vetting pipeline before Guardian sees it. Wire it as an MCP server — any role can ask for an ad-hoc scan mid-task. Configure custom rules for your own org; ride alongside the bundled 177.

The power

  • 177 rules across 13 categories — prompt injection, MCP risks, tool poisoning, GitHub Actions, secret exfiltration
  • Local-first — no SaaS, no LLM calls, no telemetry
  • Apache-2.0 — auditable, forkable, no vendor leverage
  • Cross-ecosystem package supply-chain: npm, pnpm, PyPI, Go, Rust, PHP, Ruby, Java, .NET

Why we chose it

  • Small and new on purpose — 81 stars, but a surface small enough that we read all 177 rules before shipping it. You can too.
  • The only scanner designed FOR an agent fleet, not retrofitted
  • Local-first matches our "your AI, your data" promise without an asterisk

Evolution

  • 2025 — initial release focused on prompt-injection and MCP risk rules
  • 2025-2026 — rule library expands to cover tool poisoning and Actions abuse
  • 2026 — package-supply-chain category added, cross-ecosystem coverage stabilizes

Where you might know it from

  • The agent-security scanner targeting agent-specific risks instead of generic SAST
  • The "Semgrep for AI agents" showing up in security-team Slacks in 2025-2026
  • The local-first scanner that doesn't phone home and doesn't make LLM calls
57%
fewer tokens
70%
fewer tool calls
25-35%
cheaper
Founder-published benchmark · multiple comparisons · no GPU required
CodeGraph "Your crew stops re-reading the codebase every prompt." AST INDEX · BENCHMARKED Apache-2.0 DETAILS How it lives inside RHOBEAR A background Rust worker builds the snapshot when the hub boots — done by the time you finish your coffee. Every role gets a code:lookup tool — one-answer queries instead of grep gambles. Pair with MCP — install a CodeGraph MCP server, even non-RHOBEAR agents can query it.

The power

  • AST index built with tree-sitter — exact-symbol lookups, not fuzzy similarity
  • No embeddings, no GPU, no vector database to babysit
  • Incremental: rebuilds on file change, not on every prompt
  • Founder benchmark: 57% fewer tokens, 70% fewer tool calls, 25-35% cheaper

Why we chose it

  • Crew can't be smart about a codebase they have to re-discover every prompt
  • Founder-published numbers are real, defensible, consistent
  • No GPU dependency means it runs cleanly on a laptop and a $5 VPS

Evolution

  • 2025 — initial AST-index release, tree-sitter-backed
  • 2025 H2 — PageRank-style relevance ranking added, incremental rebuilds stabilize
  • 2026 — published cross-benchmark vs embedding-based approaches

Where you might know it from

  • The structured-AST index that broke into agent-tooling Twitter on founder benchmark posts
  • The "no embeddings, no GPU" alternative to vector-search for codebase understanding
  • The thing your agent was secretly using when it stopped grepping for the wrong file

Younger than the rest — the proxy metric is the founder-published benchmark, not the GitHub star count.

Tauri "A 30MB installer a non-developer will actually double-click." ★ 107,367 Apache-2.0 + MIT github.com/tauri-apps/tauri DETAILS How it lives inside RHOBEAR Double-click the installer — Tauri shell starts, OpenClaw sidecar starts, crew on deck. No terminal. The hub UI is a Tauri webview hosting the local React app — same code on Mac, Windows, Linux. Self-host on a VPS by running just the Rust core without the webview shell.

The power

  • Rust core — fast cold boot, low RAM, signed binaries
  • ~30MB installer instead of Electron's ~300MB baseline
  • Native NSIS + MSI on Windows, .app on Mac, AppImage/.deb on Linux
  • externalBin sidecar pattern: bundle native daemons (OpenClaw) inside the installer

Why we chose it

  • The .exe has to be small enough that non-developers will actually download it
  • externalBin gives us a clean way to ship OpenClaw inside the installer
  • Apache-2.0/MIT dual license removes every distribution question before launch

Evolution

  • 2022 — 1.0 release, Electron-alternative positioning lands
  • 2023 — Mobile (iOS/Android) preview ships
  • 2024-2025 — Tauri 2.0 stabilizes, sidecar pattern matures
  • 2026 — production usage in dozens of indie + enterprise apps

Where you might know it from

  • The Rust-based desktop framework people switch to when tired of 300MB Electron downloads
  • The ~30MB shell behind a growing number of indie desktop apps in 2024-2026
  • The cargo tauri build command that made native desktop one-command-simple
Obsidian "Your second brain, in files you own — and it outlives us." 1M+ users · graph view canonized ★ 18,321 plugins + releases proprietary app github.com/obsidianmd DETAILS How it lives inside RHOBEAR Compass goals, agent transcripts, worker-session summaries get written into the vault as plain markdown. Open the Vault tab — click any node in the graph to drill into a note. Point Obsidian itself at the same folder — your second brain reads the same files the crew wrote.

The power

  • Local-first vault — every note is plain markdown on your disk. Open it in any editor. Nothing lives trapped inside our app.
  • Wiki-link backlinks build the knowledge graph automatically — no tagging chores, no folders to maintain.
  • Graph view turns your notes into something you can actually see — watch your second brain assemble itself. People don't pick Obsidian because it's efficient; they pick it because the graph is addictive.
  • The most battle-tested local-first knowledge structure there is — 1M+ users, thousands of plugins.

Why we chose it

  • Your vault outlives RHOBEAR. Rip us out tomorrow and your second brain stays — plain markdown on disk, no export step, no lock-in.
  • We built the same wiki-link + graph shape into the hub, so the crew writes into a structure you already trust.

Evolution

  • 2020 — first public beta, wiki-link + backlinks as the core idea
  • 2021-2023 — graph view and plugin API mature, community grows past 1M users
  • 2024-2026 — pattern becomes default shape for "agent-readable knowledge vaults"

Where you might know it from

  • The personal-knowledge app with the addictive graph view (2020-2024)
  • The [[wiki-link]] pattern everyone copied after Obsidian made it canonical
  • The thing your second-brain has probably been living inside for years
↓ The vault graph: 28 notes, all interlinked, autogenerated as the crew worked. ↓
Vault graph view inside RHOBEAR
← PreviousOpenClawNext →Worlds