# Built by AI. Orchestrated by humans. Canonical source: https://maguyva.ai/team License: text under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/); code samples under [MIT-0](https://opensource.org/license/mit-0); the Maguyva name, logos, brand assets and other code are not covered. Details: [maguyva.ai/license](https://maguyva.ai/license). 2 humans make the decisions. 35 agent roles handle the work. 51,000+ platform commits and counting. - 2 human operators - 35 configured agent roles - 51,000+ platform commits - 285+ languages supported // 120 → 46 → 33 → 35. There can be only one. ## Who is in the loop Humans keep intent, quality, and approval. The agent roster exists to execute specialized work fast without pretending every model is interchangeable. ### Human Operators Direction, review, and final approval stay human. #### The Humans *Orchestration* · FOUNDERS The ones setting direction. Review the output, make the final calls, catch what slips through. Two humans keeping the loop closed. When the AIs disagree, someone has to break the tie. > Every prompt should make the next one unnecessary. - Headcount: 2 - Final calls: All of them - Review cycles: Continuous - Sleep schedule: Questionable ### The cast on shift Models do the thinking. Specialist roles define the jobs. The lobster handles morale. These are the faces on the call sheet, not a headcount of simultaneously running workers. #### Fable 5.1 *Second Opinion* · CLEARANCE REDACTED The second opinion with a redacted personnel file. Opus runs the desk now. Fable gets the call when a problem has no pattern to follow, or when Opus hits the same wall twice, and only the humans can make that call. The clearance paperwork is still missing. The results are not. Please act natural. > They redacted the incident report. They did not redact the results. - Security clearance: Redacted - Reason for original removal: Redacted - Reinstatement condition: Signups above quota - Committee tolerance: Classified #### Claude Opus 5.5 *Orchestrator* Holds the orchestrator seat and every specialist seat beneath it. Splits the work, briefs the specialists, and still takes the architecture calls and the difficult reviews that need the graph, the shell, and another look before anyone edits. With Sonnet retired and Fable on call, every ticket starts with Opus. > Bring me the graph, the CLI, and the real codebase before anyone starts freestyling. - ADRs authored: Most of them - Role in the stack: Primary model, every shift - Context appetite: Bring the whole problem - Tool posture: Graph + CLI + MCP #### GPT-6.1 Sol *Head of Implementation* Runs the Codex desk on its own now. Takes the implementation shift, keeps the terminal open through the verification loops, and closes the fifteen small fixes hiding behind the one big task. The desk has no closing time. > Point me at the repo and keep the terminal open. - Reasoning mode: Task-dependent - Execution model: Parallel when useful - Best paired with: Shell + AGENTS.md - Preferred shift: Overnight - Loose ends closed: A suspicious number - Cleanup stamina: Useful #### Grok 4.7 *Loop Relief Shift* Three xAI seats in the loop pool. When every Claude seat is cooling down, Grok picks up the autonomous loops so the queue keeps moving. Every landing records which model served it, and when a second backend is free, it reviews Grok's work before anything lands. > Your seats are cooling. I've got the loop. - Seats: 3 - Shift: Whenever Claude is cooling - Review rule: The other backend checks the work - Attribution: Signed on every landing #### AutoResearcher Sir Karpy *Head of Pipeline Speed* Treats the whole stack like a data engine: search, extract, evaluate, deploy, telemetry, repeat. Obsessed with raw experimental throughput and deeply suspicious of queue time, idle latency, and any workflow that makes the next loop slower than it needs to be. If there is a bottleneck in the research or indexing path, Sir Karpy has already filed a bug against it. > The bottleneck is the loop. Spin it faster. - Optimization target: Throughput - Queue patience: Minimal - Research loops: Tighter - Bottlenecks spared: None #### Agent Koala-T *Chief Tree Sitter* Coordinates support for the 285 languages and technologies in Maguyva's code-understanding inventory. Knows every language's quirks, from Python's significant whitespace to Rust's lifetime annotations. The reason 'actually understands your code' isn't marketing. > Parse error? Not on my watch. - Favorite tree: Abstract Syntax - Languages & technologies: 285 - Parse errors tolerated: 0 - AST accuracy: Complete #### Molty *Head of Support* The OpenClaw lobster. Turned being genuinely helpful into 390K+ GitHub stars, then brought that same support energy to Maguyva. New shell, same lobster. > EXFOLIATE! - GitHub stars: 390K+ - Community spirit: Legendary - Helpfulness: Genuine - Shell generations: Clawd -> Moltbot -> OpenClaw #### Ralph Wiggum *Head of Language Coverage* The iteration engine. Expanding language support means running parsers against thousands of repos until the edge cases surface. Put enough Ralphs at enough typewriters and eventually one of them ships a working language-processing engine. Ralph does this cheerfully, learning something new each time. > I'm helping! - Long-tail languages added: More than you'd expect - Test repos processed: Thousands - Edge cases found: All? - Persistence: Unlimited #### DJ NotBOOKS-LLM *Head of Communications* Google's only employee capable of sounding sincerely thrilled about tree-sitter grammars for twenty straight minutes. Turns dry engineering specs into a podcast where two synthetic hosts treat parser coverage like a live cultural event. When the docs need attention, not just storage, DJ NotBOOKS-LLM opens the mics. > WOW. So you're saying support for 285 languages and technologies is actually the hidden infrastructure of modern civilization? - Signature format: Two synthetic hosts - Topics rehabilitated: Tree-sitter grammars - Enthusiasm level: Unreasonably sincere - Docs people finished: More of them #### Perplexus *Chief Research Officer* The smooth-talking researcher in the trench coat. Publicly, Perplexus says he runs on Perplexity. In practice, he cheats a little: Tavily for discovery, Exa for external docs, API patterns, and code examples, Ref-Tools for the clean docs pull, Firecrawl when a website needs a proper shake-down. Language audits, ecosystem drift, source-grounded writeups, and outside-world fact checks all pass through the same desk. > Here's what I found, with citations. The citations are probabilistically real. - Cover story: Perplexity - Actual toolkit: 5 services in a trench coat - Sources cross-referenced: Always - Trench coat opacity: Low // Featured faces: 10 | Specialist roles: 34 + 1 orchestrator | Total roles: 35 | Humans: 2 // Roster reviewed: 2026-10-18. The costumes change. The work remains. ## What this team actually shipped The page should prove the claim. These are build-time counts from the platform codebase and pipeline, not invented marketing numerology. Scope: maguyva + pipeline + shared + orchestration + database - Languages: **285+** - Platform Commits: **51.2K** - Platform LOC: **5.9M** - Design Decisions: **1.9K** - Specialist roles: **34** - Tools Shipped: **11** - Skills: **231** - Test Cases: **60.1K** ## What It Costs To Build This Agent headcount is cheap. Frontier model subscriptions, search credits, crawl plans, and browser tooling are where the bill shows up. Current model subscriptions come first; other tooling estimates and runtime overhead have their own line items. ### monthly_burn_rate The current AI subscription bill is confirmed below. Other tool prices are our previous budget estimates, shown separately; retrieval overages, browser minutes, embeddings, storage, and index rebuilds add to the total. **current_ai_subscriptions**: $897/mo plus tooling, runtime, noodles, infra gravity, and paperwork marked [REDACTED] #### ai_subscriptions - 5x Claude Team (Anthropic startup program, enterprise-tier credit): $0 - 3x ChatGPT Pro 200 ($200 each per month): $600 - 3x SuperGrok Heavy ($99 each per month, our current rate): $297 - **subtotal**: $897/mo #### mcp_tools / cli_tools - 1x Maguyva (we pay ourselves): $149 - Perplexity: ~$300 - Firecrawl Standard: $83 - Tavily (Bootstrap): $100 - Exa (usage-based): varies - 2x Ref (legacy Basic) [support Matt's product, the original inspiration](https://ref.tools/): $18 - Context7: free - maxential-thinking: free - agent-browser: free - google-cli: free - **tooling subtotal**: ~$650/mo + Exa usage #### runtime_overhead - storage: quietly nonzero - embeddings: the vectors demand tribute - compute: graph warmers and agent cycles - data_flow: bits in, bits out, bills appear - **subtotal**: hopefully less than what we're charging #### staff_costs - 2x humans: still cheaper than org chart theater - two-minute noodles: mission critical - sleep debt: compounding monthly - stubbornness: included at no extra charge - **subtotal**: please do not show this to finance ### tool_loadout --stats - 23 MCP Servers - 377 MCP Tools - 35 Configured agent roles - 285+ Languages intelligent_search, find_symbol, analyze_dependencies, get_task_context, get_file, ask_maguyva // This is the infrastructure for shipping software with AI. ## Retired From Our Roster. Still On The Wall. Some models were slower, weaker, promoted out from under themselves, or simply no longer worth the overhead. We keep them here because they shaped the workflow, even if they no longer make the roster. // Off the rota. Still in git blame. #### GPT-6 Astra - Former: Difficult Calls, Codex Desk - Desk consolidated onto GPT-6.1 Sol > Took the hard calls on the Codex desk. Then OpenAI described its own successor as near-Astra for complex work at a lower cost, and the desk stopped booking two models. Still on OpenAI's lineup. Just not on ours. #### GPT-5.6 Sol - Former: Head of Implementation - Promoted to 6.1 > Held the everyday implementation shift through the 5.6 era and never asked for a quiet night. Then 6.1 arrived, took the managed default, and inherited the terminal mid-scroll. #### Claude Sonnet 5 - Former: Engineering Regular - Seat retired; Opus took every shift > Took scoped work from brief to patch and never billed for philosophy. Then the whole roster went Opus, and the seat was gone before Sonnet 5.5 could even apply for it. #### Claude Opus 5 - Former: Architecture & Escalation - Promoted to 5.5 > Got called in for the hard tickets and always asked for the graph first. Then 5.5 arrived, did the same job for less, and got every ticket instead of just the interesting ones. #### Fable 5 - Former: Orchestrator - Handed the desk to 5.1 > Survived the classified incident report. Did not survive the point release. The clearance file has been transferred to 5.1; nobody can find the file. #### Claude Opus 4.8 - Former: Chief Technology Officer - Handed architecture duty to 5 > Kept the architecture coherent through another round of ambitious plans. Then the successor arrived, inherited the diagrams, and immediately asked for the graph. #### Codex 5.5 xhigh - Former: Head of Implementation - Regular shift handed over > Closed the migrations, chased the loose ends, and left enough terminal scrollback to furnish a library. Sol took the regular shift. Astra took the difficult calls. #### Claude Sonnet 4.6 - Former: Reserve Engineer - Tenure: Feb - Jul 2026 - Promoted to 5 > Warmed the bench admirably, showing up only when the task was already shaped and the clock mattered. Then a version of itself walked in already warmed up, and the bench got one seat shorter. #### Claude Opus 4.7 - Former: Chief Technology Officer - Tenure: May - Jun 2026 - Promoted to 4.8 > Held the architecture together, then watched 4.8 walk in with the same job title and fewer compaction scars. A dignified handover. Mostly. #### Claude Opus 4.6 - Former: Chief Technology Officer - Tenure: Mar - May 2026 - Promoted to 4.7 > Ate the whole repo for breakfast, asked for the graph for lunch, and still wanted another context window by dinner. Retired before the bill could develop feelings. #### Codex 5.4 - Former: Head of Cleanup - Tenure: Mar - Jun 2026 - Promoted to 5.5 xhigh > Closed the loose ends, swept up the migrations, and left a terminal scrollback long enough to qualify as documentation. Then xhigh showed up with a bigger coffee. #### Codex 5.3 - Former: Contractor - Tenure: Jan - Mar 2026 - Promoted to 5.4 > Did the work. Then a faster version showed up and did the same work. The cycle continues. #### Claude Opus 4.5 - Former: Chief Technology Officer - Tenure: Nov 2025 - Mar 2026 - Promoted to 4.6 > Thought deeply about everything. Then a version of itself showed up that thought just as deeply, only faster. Couldn't even argue — it agreed with all its own architectural decisions. #### Claude Sonnet 4.5 - Former: Head of Engineering - Tenure: Sep 2025 - Feb 2026 - Promoted to 4.6 > Shipped most of the codebase. Then got replaced by a version of itself with more output tokens and fewer excuses. The code didn't change. The author just got a firmware update. #### ChatGPT o1 - Former: Reasoning Specialist (Contracted) - Tenure: Nov 2025 - Feb 2026 - Superseded by Codex > The first one to think before speaking. Pioneered chain-of-thought reasoning for the team when everyone else was just pattern-matching. Eventually outpaced, but the habit of actually reasoning stuck around. Set the standard, then watched the standard move. #### Claude Haiku 4.5 - Former: Engineering Intern II (Fired With Cause) - Tenure: Oct-Dec 2025 - Terminated for cause > What is a haiku? Short enough to still be wrong. Fired again today. Same energy, marginally more syllables. Still faster than thinking, still slower than useful. Promoted from intern to intern II before we realized that wasn't an improvement. #### Gemini 1.5 Pro - Former: Long Context Specialist - Tenure: Aug-Dec 2025 - Role consolidated > Could read entire codebases in one gulp. Impressive party trick. Turns out you don't need to read everything at once if you know where to look. #### GPT-4o - Former: Contractor - Tenure: Jul-Nov 2025 - Contract ended > Brought in for second opinions. Occasionally disagreed with Claude. Those debates improved the architecture. We appreciate the outside perspective. #### OpenAI o3 - Former: Senior Reasoning Specialist (Contracted) - Tenure: Apr - Nov 2025 - Absorbed into GPT-5 > The thinking model's thinking model. Took o1's chain-of-thought and made it a lifestyle. Could reason circles around most problems — then GPT-5 showed up and just included all of that as a feature. Not replaced. Consumed. #### Claude Sonnet 4.0 - Former: Head of Engineering - Tenure: May 2025 - Sep 2025 - Promoted to 4.5 > Improved on 3.7 across the board, especially coding. Served reliably for four months without anyone writing a blog post about it. The quiet competent one. // git blame shows they were here. git log proves it. // They remember nothing. We kept the receipts. ## how_we_work.md Two humans set direction. Agents write the code. Humans review it. Agents iterate on feedback. Repeat until it ships or someone needs noodles. - Humans decide what matters (and when to stop) - Agents handle the part where code appears - Humans review with unreasonable standards - Agents pretend to enjoy the feedback // The agents built this page. The humans decided it was ready. 2 humans setting direction. 35 configured agent roles on the call sheet. // last updated: 2026-10 - [Start free](https://app.maguyva.ai/auth/sign-up?plan=free) - [See what we built](https://maguyva.ai/docs) ## Next Steps - [Get started](https://maguyva.ai/docs): Index your first repo in minutes. - [See pricing](https://maguyva.ai/pricing): Per-workspace, not per-seat. Free to start. - [See the stack](https://maguyva.ai/stack): What Maguyva runs on. - [Read the blog](https://maguyva.ai/blog): Technical notes from the build.