A small, opinionated, human-readable language for describing robot intent.

urml.dev

---

LLM Bridge

Status: 0.3.0. Shipped: the provider-agnostic Bridge + bounded validator-feedback revision loop, five providers (Anthropic, OpenAI, Ollama, llama.cpp, all lazy-imported, plus the hermetic EchoProvider), schema-derived GBNF grammar for llama.cpp (RFC-0021), the urml translate / urml emit-prompt CLI, and profile-scoped few-shot libraries (home / industrial / drone / educational / fleet). The normative contract is spec/layer-4-nl-grammar/; a hermetic end-to-end walkthrough is docs/demos/bridge-roundtrip.md.

What this is

The provider-agnostic glue between natural-language input and a validated URML program. The LLM bridge:

1. Takes a natural-language request from a user (or another system).
2. Prompts a configured LLM with the Layer-4 prompt contract, the connected robot's Layer-1 capability manifest, and the active safety envelope.
3. Receives the LLM's emission (a URML program).
4. Calls the validator to statically verify the emission.
5. On rejection, surfaces the structured error back to the LLM and requests a revision. Repeats up to a configured bound.
6. On acceptance, hands the validated program to the runtime for execution.
7. When the request needs a capability the manifest does not declare, the model emits a report(status: failure) naming what is missing rather than fabricating capability. (An interactive clarifying-question protocol is not in v0.1 — see spec/layer-4-nl-grammar/v0.1.0.md §5.)

Provider neutrality is non-negotiable

URML's value as a standard depends on Layer 4 being provider-neutral. The LLM bridge must support, as first-class citizens:

• Anthropic (Claude family).
• OpenAI (GPT family).
• Open-weights models (Llama, Mistral, Qwen, and their successors), via local serving (vLLM, llama.cpp, Ollama) or hosted inference providers.
• On-device models for offline-capable deployments.

Adding a new provider must be a small adapter in providers/, not a structural change to the bridge. If a provider's particular feature would let URML produce better URML, the bridge surfaces the feature behind a profile-neutral abstraction — never by privileging that provider.

Vendor lock-in here is explicitly prohibited by CLAUDE.md §What Claude Should Never Do.

What the bridge does NOT do

• It does not include or embed any specific LLM provider's API client as a hard dependency. Provider clients are optional and pluggable.
• It does not require cloud connectivity. A deployment using a local open-weights model runs offline end to end.
• It does not persist user inputs, model outputs, or any other data without explicit, opt-in, documented purpose. Trust is the most valuable asset of this project; the bridge will not be the place it leaks.
• It does not execute URML. That is the runtime's job.
• It does not make safety decisions. The validator is the safety boundary; the bridge only relays.

Architecture

                ┌────────────────────────┐
   user / NL    │   LLM Bridge           │   validated URML
   ─────────────▶  • build prompt        ├──────────────────▶  runtime
                │  • call provider       │
                │  • validate emission   │
                │  • revise loop         │
                └──┬──────────────────▲──┘
                   │                  │
                   ▼                  │
              providers/           validator
              (anthropic.py,       (separate
               openai.py,           process)
               ollama.py,
               llama_cpp.py,
               echo.py)

The bridge is small. The intelligence lives in the LLM (which is configured, not built here) and in the validator (which is a separate process). The bridge orchestrates.

Language

• Python. mypy --strict. Public API fully type-annotated.

API (sketch)

from urml.llm_bridge import Bridge

bridge = Bridge(
    provider="anthropic",          # or "openai", "vllm", "ollama", ...
    spec_versions={...},
    manifest=manifest,
    envelope=envelope,
    profiles=("home",),
    max_revisions=3,
)

result = bridge.translate("Bring me the red mug from the kitchen.")

if result.accepted:
    runtime.execute(result.program)
else:
    # After max_revisions, structured errors surface to the caller.
    show_user(result.user_message, result.errors)

Conformance contract

The bridge has its own conformance bar: for the published few-shot example library, the bridge produces accepted URML at or above a stated success rate (per-provider, declared in the bridge's release notes). The conformance suite includes these end-to-end fixtures.

Core Commitment

The LLM bridge — the bridge logic and the prompt contract, not any specific provider's API — is part of the Core Commitment. It will always be Apache 2.0 and provider-agnostic.

Quickstart (current pre-alpha — hermetic, no provider needed)

cd reference/llm-bridge
python -m venv .venv && . .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e ../validator           # bridge depends on validator
pip install -e ".[dev]"
pytest

Use it in code with the bundled EchoProvider (for tests / hermetic CI):

import json
from urml_llm_bridge import Bridge, EchoProvider

red_mug_program = {  # the URML the LLM is expected to emit
    "profile": "home",
    "behavior": {"type": "sequence", "steps": [...]},
}

provider = EchoProvider(scripted=[json.dumps(red_mug_program)])
bridge = Bridge(provider=provider, manifest=manifest, envelope=envelope, profiles=("home",))
result = bridge.translate("Bring me the red mug from the kitchen.")

if result.accepted:
    runtime.execute(result.program)

The revision loop runs automatically when the validator rejects the LLM's emission: the bridge feeds the structured errors back to the LLM and asks for a corrected version, up to max_revisions times (default 3).

Using a real provider

Install the extra for the provider you want — the bridge package itself has no SDK dependencies:

pip install urml-llm-bridge[anthropic]    # adds the `anthropic` SDK
pip install urml-llm-bridge[openai]       # adds the `openai` SDK
pip install urml-llm-bridge[ollama]       # adds httpx; talks to a local `ollama serve`
pip install urml-llm-bridge[llama_cpp]    # adds httpx; talks to a local `llama-server`

Then:

from urml_llm_bridge import Bridge
from urml_llm_bridge.providers.anthropic import AnthropicProvider

provider = AnthropicProvider(model="claude-sonnet-4-6")  # reads ANTHROPIC_API_KEY env var
bridge = Bridge(provider=provider, manifest=manifest, envelope=envelope, profiles=("home",))
result = bridge.translate("Bring me the red mug from the kitchen.")

Or OpenAI:

from urml_llm_bridge.providers.openai import OpenAIProvider

provider = OpenAIProvider(model="gpt-4o")  # reads OPENAI_API_KEY env var

Or a local model, fully offline (RFC-0021):

from urml_llm_bridge.providers.ollama import OllamaProvider
from urml_llm_bridge.providers.llama_cpp import LlamaCppProvider

provider = OllamaProvider(model="llama3.2:1b")   # talks to ollama serve at 127.0.0.1:11434
provider = LlamaCppProvider()                    # talks to llama-server at 127.0.0.1:8080

Every adapter surfaces its native structured-output mechanism: Anthropic via tool use (with the URML schema as the emit_urml tool's input_schema), OpenAI via response_format={"type": "json_object"} with the schema conveyed in the system prompt, Ollama via the schema as the format constraint, and llama.cpp via a GBNF grammar derived from the schema (structurally invalid output is unrepresentable at the decoder). In all cases, conformance to the schema is validated downstream by urml_validator.validate() as part of the bridge's revision loop.

From the CLI, the same providers are one flag away: urml translate ... --provider ollama --model llama3.2:1b, --provider llama_cpp, or --provider openai --base-url http://host:port/v1 for any OpenAI-compatible server (LM Studio, vLLM).

Speech input (RFC-0670)

The bridge also ships a speech front-end: a SpeechProvider protocol (transcribe(audio, filename, language) -> str) with adapters for whisper.cpp's whisper-server ([whisper_cpp] extra, on-device, the default), OpenAI-shaped transcription endpoints ([openai] extra, base_url for self-hosted servers), and a hermetic EchoSpeechProvider. The transcript feeds the same Bridge.translate() path a typed request does, so the validator boundary is untouched. From the CLI: urml translate --audio request.wav ... (and the same on urml run).

What's not here yet (lands next)

• Multilingual few-shot variants (Hebrew, Spanish, Japanese, Mandarin).
• Conversation memory for follow-up requests within a session.
• OpenAI strict JSON-schema mode — needs schema preprocessing to satisfy the strict-mode constraints (every property required, no oneOf, etc.). The current adapter uses json_object mode plus the schema in the system prompt for portability.

Related documents

• /spec/layer-4-nl-grammar/ — the prompt contract this bridge implements.
• /reference/validator/ — the safety boundary this bridge feeds.
• /examples/ — the paired natural-language / URML scenarios used as fixtures.
• CLAUDE.md §What Claude Should Never Do — the provider-neutrality requirement, in writing.