Willow IntelligenceWillow IntelligenceRequest Developer Preview
Developer Preview · willow.core

One continuity signal.
Every surface. Every provider.

Willow Core exposes the same coherence schema through REST, Python SDK, and MCP — thread integrity, fragmentation risk, recommended action, intent vector, and a compact continuity artifact.

SDK Quickstart Schema Reference MCP Tools Provider Matrix
Architecture

App → Willow Core → Providers.

Willow sits inline between your application and the model provider. It observes every request, scores coherence, and emits a recommended action before the response reaches your user.

APP LAYERChat UIexternalAgentexternalWorkflowexternalWILLOW COREAnchorwillow.core/v1Observe → Scorewillow.core/v1Act → Stabilizewillow.core/v1PROVIDERSOpenAIexternalAnthropicexternalMistral / OSSexternal← CoherenceSignal (thread_integrity, fragmentation_risk, recommended_action)
Schema

CoherenceSignal

Example response. Hover any key to inspect its technical definition.

POST /v1/coherence · example response
01{
02  "thread_id": "thr_8f3a...",
03  "thread_integrity_score": 84.186,
04  "fragmentation_risk": "watch",
05  "recommended_action": "verify",
06  "intent_vector": { ... },
07  "continuity_artifact": "cont_a91f...",
08  "provider": "openai/gpt-4.1"
09}
Surfaces

API · MCP · SDK

REST
v1

HTTP API

  • POST /v1/coherence
  • GET /v1/threads/:id
  • POST /v1/anchor

Bearer auth. JSON. Streaming via SSE.

MCP
v1

Model Context Protocol

  • willow.observe
  • willow.score
  • willow.reanchor
  • willow.verify
  • willow.replay

Drop-in for any MCP-aware agent runtime.

SDK
v1

Python · Typed client

  • willow.Client(api_key=...)
  • client.coherence(thread, output)
  • client.stream(thread)

Async + sync. Pydantic models. Retry semantics built-in.

Quickstart

Python · ~6 lines

bash
01# install
02pip install willow-core
03
04# authenticate
05export "WILLOW_API_KEY"="wk_live_..."
python · score one turn
01from "willow" import Client
02
03client = Client()
04
05signal = client.coherence(
06    "thread_id"="thr_8f3a",
07    "anchor"="draft GDPR clause, EU only",
08    "output"=model_output,
09)
10
11"if" signal.recommended_action == "reanchor":
12    "raise" DriftDetected(signal)
MCP

Five tools for agent runtimes.

tooldescription
willow.observeStream a thread under observation. Returns CoherenceSignal events.
willow.scoreScore a single turn against the current anchor.
willow.reanchorRe-establish the anchor with verified context.
willow.verifyRun a verification pass on a flagged output.
willow.replayReplay a thread with full signal trace for inspection.
Early validation

Careful signals, not benchmarks.

SignalResultContext
Live Core smoke — API · SDK · MCP · multi-model pathpassedcontrolled launch smoke
CoherenceSignal fields returned in live responses4thread_integrity_score · fragmentation_risk · recommended_action · intent_vector
Provider-adaptive controls generated by Willowconsistentfocused provider validation
Continuity artifact vs. source transcript~14.6%synthetic long-context compression test
Gemini provider-adaptive followed-rate37.5% → 62.5%focused validation; high-load/urgent reached 100% after policy refinement
Automated tests passing63launch smoke / test run
Controlled private tests · should be validated in customer environments · results depend on workload, provider, and integration.
Honest limits

What we don't ship.

Cold-start behavior
First signal on a fresh thread incurs additional anchor compute. Subsequent turns are lighter; exact latency depends on provider, payload, and region.
Provider scope
Streaming SSE supported on OpenAI and Anthropic during preview. Mistral and OSS are request/response only in v1.
Region
us-east-1 during Developer Preview. Additional regions planned.
Rate limits
Preview: 60 rps per key. GA: configurable.