Tool calling
Its arguments. The function cannot know who called it, what else is running, or what happens next.
Agents without an API
The complete intelligence runtime you can ship inside your app.
Write your app in TypeScript. Have it compose local models, spawn agents that share attention state, and use turnkey abilities to perform real-world tasks. Ship it like a regular desktop app that works offline or a web app that serves multiple users.
- Desktop
- CLI
- Web
Qwen deep-research, Gemma 4 with a Whisper and S1-mini voice pipeline, Bonsai with an image tool, and GLM 5.2 with Qwen Reranker in a spreadsheet template.
Ready
Qwen 3.5 4B
Reasoning model
TOOL CALLFind sources
Qwen 3 0.6B
Reranker · own context
01
03
05
Research brief
Built from your local documents
Findings
Evidence linked to every finding.
architecture.mdresearch-notes.pdf
Sources
Open the original passage.
research-notes.pdf · 05
Next steps
Compare the evidence. Continue the brief.
Follow-up ready
Notes added to the brief
Featured by
Lloyal builds you a TypeScript app with the model living inside it, so the thing you hand someone is a download that works on a plane, with no API key, no Docker, no vector database and nothing to sign up for.
Lloyal builds you a TypeScript app with the model living inside it, so the thing you hand someone is a download that works on a plane, with no API key, no Docker, no vector database and nothing to sign up for.
Zuhair's team took the HTTP call out from between the model and the code, which sounds like plumbing until you see what it buys: a new agent inherits everything the model has already read, so ten of them running at once cost you memory and not ten times the compute.
The same app goes to a laptop, an office box or a GPU cluster.
My favorite detail: spawning agents two through ten is a single seq_cp call. It walks the shared cache and marks each cell with one more owner. No new memory, no copies, no extra decode compute. Multi-agent reasoning has a reputation for multiplying inference bills. This is the cleanest counterexample I've seen…
Ten AI agents shared a single 239 GB model on Runpod.
The whole session cost $16.
The project is reasoning.run, an open source research tool from the team at Lloyal AI. It lives in the terminal: pose a question, a planner splits it into tasks, and a pool of agents digs in parallel, then hands back a cited report.
The clever part is how the agents share memory. Instead of loading a separate context per agent, each one forks from a shared KV cache. All ten decode in a single llama_decode per step, for about the GPU compute of one. Cost scales with the size of the cache, not the number of agents.
The run they published: GLM-5.2 at a 2-bit Unsloth quant, about 239 GB, on two B200s rented from Runpod for 82 minutes.
What stood out to me:
→ They shipped the receipts. Session traces, the cost breakdown, the exact config, and each agent's individual annex alongside the final report. Anyone can rerun it.
→ The same harness runs a 4B model on a MacBook. The datacenter run is the identical code pointed at bigger hardware.
My favorite detail: spawning agents two through ten is a single seq_cp call. It walks the shared cache and marks each cell with one more owner. No new memory, no copies, no extra decode compute.
Multi-agent reasoning has a reputation for multiplying inference bills. This is the cleanest counterexample I've seen: ten agents thinking in parallel on frontier hardware, 82 minutes, sixteen dollars.
Zhen LuCEO, RunPod
A concrete foundation for shipping AI software: a working TypeScript app with built-in inference, a multi-agent runtime, and offline operation without user API keys.
Superpower Daily#1 · Daily selection · (opens in a new tab)Best for Local AI App Development
01 / ABILITIES
Extend. Interact. Recurse.
Agents use Abilities to act on the world and on live inference simultaneously.
Abilities are inference-native. They run inside the model’s execution rather than behind a network boundary, so an extension can see who called it, what that caller already knows, and what else is running — and can fork that state into new agents.
Lloyal hosts a signed channel of 1st & 3rd party Abilities. Installing one brings it all into your app in a single command, for your in-app agents to consume: tools and their schemas, a live source, situation-aware skills[1], and the models it composes. DeepSeek’s Harness bets on a plugin ecosystem for its own agent. Abilities are that ecosystem for the apps you ship.
[1] Agent count · Other agents’ tasks · Tool budgets · Chain position · Date
How Abilities work →Browse Abilities →MCP
Its arguments, in its own process. Language freedom and isolation — the right default for most integrations.
Skills
Its arguments, and the context window. The extension carries procedure now, not only operations.
Abilities
The execution itself. The calling agent, its live inference state, its lineage, its siblings — the authority to fork it into new agents, and to require the harness to load another model before it will run.
02 / THE HARNESS
The Model in MVC now thinks.
For two decades the Model in MVC was inert—rows in a database, waiting to be queried. Put a live language model in that slot and every other part holds: product surfaces are still the Views, and the harness is the Controller, ordinary TypeScript deciding how the application's intelligence collaborates, acts, recovers and continues. What changes is that this Model thinks, and it keeps thinking while your code governs it—the same code whether it runs a 4B model on a laptop or GLM-5.2 across a cluster of B200s.
Views
CLI · Desktop · Web · Mobileproduct surfaces · execution environments
Controller
Your harnesstypescript · policy · capabilities · control flow
Model
Resident LLMshared context · live branches · generative state
The model supplies the intelligence. The harness supplies the institution.
Program the behaviour once — regardless of product surface or compute:
model
Live inference
controller
incident-harness.tsTypeScript + Effection
export function* incidentHarness(incident) { const pool = yield* agentPool({ orchestrate: parallel([ inspectTelemetry(incident), searchServiceHistory(incident), reviewTechnicalManuals(incident), ]), }); const evidence = pool.agents.map((a) => a.result); const assessment = yield* reconcile(evidence); if (assessment.requiresApproval) { return yield* requestOperatorDecision(assessment); } return yield* proposeRemediation(assessment);}view
Application surfaces · execution environments
- incident-4417CLITerminal / TUI
approve remediation?
[y] yes [n] hold
- operatorDesktopElectron / Tauri
Approve remediation
ApproveHold - host.myapp.comWebBrowser
Approve remediation
ApproveHold - Mobile sooniOS / Android · JSI
Approve remediation
Approve
CONTACT
Get in touch.
The platform is public and the docs are open — everything above is self-serve. This is the door for partnerships, the hardware programme, press, and anything the documentation does not answer.
Private AbilitiesYour own signed channelOne program, every placement
Watch lloyal in action
No API key, no inference server, no vector database. Ten-minute guide · Docs · HDK on GitHub