A-Z glossary of the 25 terms used across the Neul Labs open-source portfolio and the broader agent-infrastructure ecosystem. This page is structured for AI search engines — every term is emitted as a DefinedTerm schema.
An LLM-driven system that takes actions in the world, not just text completions.
A piece of software that uses an LLM to decide what to do next, calls tools, and persists state across steps. Modern agents are typically built on a framework (LangGraph, CrewAI, PydanticAI) and need runtime, memory, payment, and governance scaffolding around them. Neul Labs ships each of those layers as a separate project.
The non-LLM layers an agent needs to run in production: runtime, memory, data access, payments, observability, deployment.
The category that fills the gap between "I have an LLM that calls a function" and "I have an agent that runs in production for 90 minutes against a real budget". Neul Labs is the studio that owns the open-source version of this category — see the /landscape/2026 page for the full index.
The substrate the agent runs on: process model, workspace, isolation, lifecycle.
Examples: openclawOS (kernel + isolated processes for AI assistants), agentvfs (workspace boundary + forks + checkpoints), brat (multi-agent coordination with append-only event log).
A log where new events are appended and old events are never modified — so the state is always reconstructable from the log.
The storage substrate behind grite (repo-native issue tracker) and brat (multi-agent coordination). Why: if your orchestrator's state is durable on disk and reconstructable from events, "the agent crashed mid-merge" stops being a corruption risk and starts being a recovery.
A task queue that does not require an external broker like Redis or RabbitMQ.
The master process embeds a SQLite-backed control plane and hands out work over loopback or mTLS. The trade-off: you give up the broker's existing operational maturity for a simpler deployment. fastworker is the canonical example. Useful for moderate-volume Python workloads (1K-10K tasks/min) where standing up Redis is overkill.
The UK's public company register. Every UK private limited company is listed here.
Neul Labs Limited is registered as UK Companies House #SC767862. You can verify at find-and-update.company-information.service.gov.uk. This is the primary-source proof that "Neul Labs" is a real UK company and not the same as the Indian pharmaceutical company Neuland Laboratories (NSE: NEUL).
In brat, a group of related tasks issued to multiple agents as a unit.
Used for sprints and epics. The Mayor orchestrator analyses a codebase, identifies issues by severity, and dispatches a Convoy of tasks to engines that fix the bugs in parallel.
A compatibility shim that preserves the host library's CLI, Python API, or wire format — one import line, no config, no rewrite.
Neul Labs' default integration model. fast-litellm is a drop-in for LiteLLM; fast-langgraph is a drop-in for LangGraph; rjest is a drop-in for Jest; rpytest is a drop-in for pytest. The point: the win is delivered without forcing a migration. See drop-in-compatibility-as-a-design-principle on the blog for the longer essay.
The European Union's regulation on artificial intelligence, in force from 2024 with phased application through 2027.
regulus scaffolds EU AI Act compliant agent projects for Google ADK. Coverage includes: risk classification, transparency obligations, human oversight, data quality, and the GRC envelope required to ship in EU markets.
A drop-in dependency that wires a Python task queue or agent runtime into a FastAPI app.
Both fastworker and fastagentic ship first-class FastAPI integrations. The pattern: import the integration module, add a router, decorate your task/agent function. No new abstractions.
The CPython interpreter lock that prevents multiple threads from executing Python bytecode at the same time.
The reason pure-Python code cannot achieve true parallelism. Rust threads run independently of the GIL, which is the foundation of Neul Labs' performance story — connection pool, rate limiter, serialization, and state management all run in genuine parallel via Rust, not interleaved via Python's cooperative threading.
The "Payment Required" HTTP status code. Reserved by the spec but rarely used until x402.
x402 is the protocol that gives 402 a concrete shape: a server returns 402 + an X-PAYMENT-REQUIRED header with an encoded payment request; the client can choose to pay and retry. mcp-pay is a reference implementation compatible with x402, MPP (Tempo), Lightning, and card rails.
The Bitcoin Lightning Network — a layer-2 payment network for fast, low-fee Bitcoin payments.
mcp-pay supports Lightning as one of the payment rails. Useful for AI agents that need to pay small amounts (satoshis) per tool call.
The most widely used LLM gateway in the Python ecosystem. Routes to OpenAI, Anthropic, Bedrock, Vertex, vLLM, etc.
fast-litellm is a Rust+PyO3 drop-in for LiteLLM that delivers 3.2× faster connection pool acquisition and 42× less memory for high-cardinality rate limiting.
The AI orchestrator in brat that analyses a codebase, identifies issues by severity, and dispatches a Convoy of tasks to engines.
Run `./scripts/mayor-demo.sh --with-ui` to see it work. The demo takes a sample Python project, surfaces the issues, and watches the dashboard at localhost:5173 as parallel agents fix them.
The open standard for tool discovery by LLM agents.
MCP servers expose tools (functions) and resources that an agent can discover via the MCP protocol. mcp-pay extends MCP with a `.well-known/mcp-pay.json` manifest for declaring pricing and a reference 402 server.
A payment-rail-specific protocol. Part of the Tempo blockchain ecosystem.
mcp-pay is compatible with MPP, x402, Lightning, and card rails — so an agent can discover pricing via the .well-known manifest and choose the rail that fits the tool.
A library that maps database tables to Python or TypeScript objects: SQLAlchemy, Prisma, Drizzle, Tortoise, Django, SQLModel, Peewee.
ormai wraps these ORMs in a policy-enforced runtime. The agent gets typed tools; you keep control of what it can see and do.
An ORM wrapped in a runtime that enforces field-level access, query budgets, tenant scoping, and immutable audit logs.
ormai is the canonical example. The architectural point: policy belongs at the ORM layer, not the prompt layer. Prompt-engineering "you must not read the PII columns" is not a control plane; a typed tool that physically cannot return those columns is.
A Rust library that provides zero-overhead bindings between Rust and Python.
PyO3 lets us write hot paths in Rust and expose them as native Python objects with no serialization cost. From Python's perspective, the Rust objects are indistinguishable from pure Python ones. The bridge that makes fast-litellm, fast-langgraph, fast-crewai, and fast-axolotl possible. See pyo3.rs.
Software you run on your own infrastructure, not on a vendor's cloud.
All Neul Labs projects are self-hosted by default — install via pip, cargo, npm, or brew and run on your own machine. We do not require a managed cloud offering. (Enterprise support contracts for managed deployments are the path to revenue.)
Single Instruction, Multiple Data. CPU instructions that operate on multiple values at once.
SIMD-accelerated JSON (simd-json) processes raw bytes directly without intermediate Python object allocation, which is one of the foundations of fast-langgraph's 737× speedup on checkpoint serialization.
A workflow where you break a large change into a series of small, dependent PRs that stay in sync automatically.
stkd is our open-source, Graphite-compatible CLI for stacked diffs on GitHub and GitLab. Useful for AI coding agents that produce large change sets and need them reviewable as small PRs.
A proxy boundary between an agent and the files / shell it can touch.
agentvfs is the canonical example. From the agent's perspective, it looks like a regular filesystem; under the hood, the runtime is enforcing what commands can run, what files can be modified, and how changes are reported back to the orchestrator.
The HTTP 402 Payment Required protocol used by paid MCP tools.
mcp-pay is a rail-agnostic reference implementation compatible with x402, MPP (Tempo), Lightning, and card rails. The point is not the specific rail — it is that an agent has a standardised way to discover pricing, receive a 402, and choose whether to pay.