MaiiaM AgentStudio + AnythingLLM-AAMT
The first developer stack where small local models reason like frontier ones — policy-gated, ledger-attested, replayable. Hover any node to see how the layers connect.
The cognitive core. Decompose intent, plan, execute, assemble — every step validated against the AAMT mathematical foundations. Turns Llama-3.1-8B-class local models into rigorous multi-step reasoners.
Three surfaces. One reasoning IP.
A self-calibration suite: author the tooling in AgentStudio, build with it in Thoughts, ground every step in your own corpus through AnythingLLM-AAMT. Wings reasoning is the common thread.
AgentStudio
The author surface
Compose agents, instruments, and reasoning workflows on a visual canvas. The IDE-native VSIX lives where you already code. AgentStudio is where self-calibration tooling is built.
You are hereThoughts
The build-with surface · thoughts that think back
A conversational shell for spinning up customware on the fly — twin-inference UI, skill + memory bridge into Wings Reasoner, and a marketplace of community customware. The umbrella for everyday self-calibration.
Explore ThoughtsAnythingLLM-AAMT
Your library, your reasoning
The corpus fabric: workspaces, embeddings, RAG, MCP host. The sovereign substrate that AgentStudio authors against and Thoughts builds on. Public OSS at the substrate, private modules loaded over a clean process boundary.
Open the corpus fabricThree Layers. One Reasoning System.
Each layer ships independently. Each one is the strongest in its category. Together they collapse the gap between local models and frontier ones.
Wings Reasoning
AAMT-grounded inference loop
The cognitive core. Wings turns small local models (Llama-3.1-8B class) into rigorous multi-step reasoners by validating every step against the AAMT mathematical foundations. Decompose intent, plan, execute, assemble — with a structured AAMT trace at every layer.
MaiiaM AgentStudio
Multi-agent orchestration UI
Author, compose, and ship agents through a visual canvas — 78 AAMT instruments, AnythingLLM workspace nodes, and Wings reasoning blocks in one drag-drop surface. Ships as a standalone web app AND a VS Code/Cursor/Antigravity extension that lives where you already code.
AnythingLLM-AAMT
Sovereign substrate, MIT-licensed fork
Workspaces, embeddings, vector stores, RAG, chat UI, MCP host — runs entirely offline with no DNS, no egress required. The public substrate is fully open-source; AAMT/Wings/CTK load over a clean process boundary so proprietary IP is never vendored into the OSS tree.
The wiring beneath the three faces.
Wings is the reasoning IP. AAMT-MCP is how agents act. The Reasoning Provider API is how the fabric stays open. Every face of the suite speaks the same protocols.
Wings Reasoner
The cognitive core. Decompose · Plan · Execute · Assemble, validated against the AAMT mathematical foundations. Exposed as a standard MCP server so AgentStudio, Thoughts, and AnythingLLM-AAMT all consume it the same way.
- Skill bridge — wings-reasoner skills are first-class in customware
- Memory bridge — vault artifacts feed Wings context
- Twin-inference UI hook — F+M passes surface in the chat surface
AAMT-MCP
The workspace operations server inside the AnythingLLM-AAMT repo. File · git · npm · scan · feature-map. Lets agents author and ship code from the canvas while staying inside CTK policy.
- 1,221 LOC of typed MCP tools
- Loaded by AnythingLLM-AAMT and AgentStudio VSIX alike
- CTK-policy gated at every tool call
Reasoning Provider API
The pluggable contract that lets any reasoning engine register with AnythingLLM-AAMT — Wings is the reference implementation. Drop in a new provider and every face of the suite picks it up.
- Documented in REASONING_PROVIDER_API.md
- Provider registration · trace emission · policy callbacks
- Lets third parties extend the stack without forking
A durable record of how you calibrate.
The vault holds the artifacts of your self-calibration practice — readings, journals, customware, citations — each signed with a workspace key so lineage is provable. Backed by packages/netergaiiam/vault.
Readings & oracles
Every Oracle session, every astrological calibration — written to the vault as a signed artifact you can return to, diff against, and replay.
Journals & reflections
Self-calibration work product: journal entries, reflection threads, and the breath-state context they were captured in.
Customware manifests
The Thoughts customware you build is stored here as a portable manifest — instruments, prompts, gates, reasoning hooks, and the library citations it grounds against.
Library citations
Every AnythingLLM-AAMT workspace pull, every wisdom-matcher hit — bound to its source so the agent's reasoning chain is fully traceable to your own corpus.
Build agents that actually reason.
Composable, governed, ledger-attested reasoning — from the editor to the canvas to the deployed workspace skill.
Multi-agent orchestration
Author multi-agent workflows visually. Sessions, relays, quality gates, token budgets — all first-class. One canvas spans Claude, Gemini, GLM, and any local model exposed via Ollama or LM Studio.
VS Code / Cursor / Antigravity native
VSIX extension v15. Lives in the IDE you already use. Spawn agents, inspect reasoning traces, run workspace ops — without leaving the editor.
Reasoning Search
Every Wings reasoning pass produces a structured, queryable trace. Ask: "show me every conclusion we reached about vendor X using AAMT layer 3 in Q2." Compliance gold; nobody else has this.
Reasoning Replay & Diff
Run yesterday's trace against today's corpus to detect knowledge-base drift. Compare two model traces step-by-step — prove that small models match big models structurally, not just lexically.
78 AAMT instruments
MaiiaM Canvas exposes 78 reasoning instruments — every TERA dimension, breath operator, sign-inversion layer — plus AnythingLLM workspace nodes and Wings blocks. Drag-drop authoring, compiles to a portable manifest.
MCP-first architecture
Wings exposes itself as an MCP server. AnythingLLM hosts it natively. The aamt-mcp workspace ops server (1,221 LOC) handles file/git/npm/scan/feature-map operations — extensible from your editor or terminal.
Local-model reasoning that actually works
Wings turns Llama-3.1-8B into a reliable multi-step reasoner. Goal: AAMT-grounded local model beats vanilla local model by ≥30% on public eval suites and reaches ≥80% of GPT-4-class performance on reasoning tasks — entirely offline.
Reasoning Unit Tests
TDD for thinking. Assert that for prompt class X, the agent must invoke layer N and consult workspace Y. Lock down emergent agent behavior the same way you lock down code.
One-click workspace skill deploy
Compose an agent in AgentStudio over your AnythingLLM workspace data, deploy it as a workspace skill — your team uses it from chat, voice, or the IDE without ever leaving the surface they're already in.
Sovereign by design. Provable by default.
Built for the buyers who can't choose between AI capability and data sovereignty. Government, defense, healthcare, legal, financial — your air gap stays intact.
Air-gapped deployment
Reference deployment runs fully offline — no DNS resolution, no egress, no callbacks. The first credible sovereign-AI stack for regulated environments.
Cryptographic attestation
Every reasoning trace is signed with a workspace key. Tamper-proof audit log. Defense, healthcare, and legal teams can finally say yes to AI assistance.
Policy-bounded execution
CTK enforces OPA/Rego policies and YAML templates at every tool call. No tool call permitted outside policy. No exception. Provable governance, not promised.
Zero-egress reasoning
AAMT/Wings/CTK load over a clean process boundary; the public AnythingLLM fork never touches proprietary IP. Your private corpora stay on your hardware. Your model weights stay on your hardware.
Reasoning Ledger
Every Wings pass produces a queryable, replayable, signable artifact. Compliance review answers questions like "what corpus did the agent consult before issuing this recommendation?" — definitively.
HIPAA / SOC 2 / FedRAMP-aligned
Architectural choices align with the controls regulated buyers actually care about — workspace isolation, key-scoped access, signed action logs, deterministic policy enforcement, no third-party data egress.
Three deployment modes. One toggle.
Cloud by default, hybrid by preference, local-only when the data can't leave. Switch in the Messenger settings — server-validated tier gate, no client-side spoofing.
Cloud
Default · fastest setup
- Hosted Claude / Gemini / GLM via the platform
- Zero local setup — works in any browser
- Fast onboarding, predictable latency
- For most users getting started
Hybrid
Local first · cloud fallback
- Try local first — fall back to cloud only when needed
- Privacy when possible, reliability when required
- For developers iterating · mixed workloads · CI runs
- Auto-detects Ollama, LM Studio, olmx, AgentStudio MCP
Local Only
Air-gapped · zero data egress
- All inference runs on your hardware
- No DNS, no egress — your prompts never leave the device
- Hard refuse if local runtime unreachable (no silent cloud fallback)
- For regulated industries · sovereign deployments · classified workloads
AgentStudio Cloud is tier-gated. Cloud access is reserved for AgentStudio Developer, Practitioner+, Lifetime, and Admin tiers. Seeker and Companion run local-first only. Tier checks are enforced server-side against MemberPress — no client-side spoofing.
Watch an agent think.
Every Wings reasoning pass produces a structured Odu trace. Hover any step to see the agent's archetype state, breath shell, and intent at that moment. Sign it, replay it, diff it, query it — reasoning becomes infrastructure.
Sample trace · 6 steps · audit query · hover the step columns to inspect
The corpus fabric has its own page now.
The sovereign AnythingLLM fork — workspaces, embeddings, RAG, MCP host, reasoning provider plug-in — is the substrate AgentStudio and Thoughts both build on. Full pitch, architecture deep-dive, and deployment shapes live on the dedicated page.
Three artifacts. One ecosystem.
Available now to active beta members. Each builds the same reasoning stack — pick the surface that fits your workflow.
MaiiaM AgentStudio
Web app + desktop build
Multi-agent orchestration UI. 78 AAMT instruments, visual canvas, sessions, reasoning traces. Runs in the browser or as a packaged desktop app for offline workflows.
AgentStudio VSIX
VS Code · Cursor · Antigravity
The same AgentStudio capability inside the IDE you already use. 20 agents, 36 skills, custom MCP server. Install via VSIX from the Extensions panel.
AnythingLLM-AAMT
Public OSS substrate + private modules loader
MIT-licensed AnythingLLM fork preconfigured to load Wings, AAMT Foundations, and CTK at runtime. Single docker-compose for the air-gapped reference deployment.
Releases signed with the AsAManThinks publisher key. Verify your download before installing — see signing & verification.
The reasoning stack frontier labs are spending billions to build.
We made it open at the substrate, sovereign at the deployment, and mathematically rigorous at the core. It runs on your hardware, follows your policy, and lets the smallest model in your toolkit out-think the most expensive cloud you've considered.