DeepTrail
Most AI research tools optimize for producing an answer. DeepTrail optimizes for making the path to that answer inspectable.
DeepTrail is a local-first research and decision workspace where a human and a WebMCP-aware browser agent work against the same visible state. Instead of burying reasoning inside chat history, DeepTrail structures an investigation into questions, sources, claims, evidence relationships, counterarguments, confidence changes, research gaps, comparisons, and decisions.
Why WebMCP is essential
DeepTrail needs the human and agent to operate on the same application state. Copy/paste or generic DOM automation would lose stable entity IDs, provenance relationships, tool intent, attribution, and the distinction between supporting, contradicting, and qualifying evidence.
The site exposes explicit research actions through WebMCP. A browser agent can inspect the active workspace, identify unresolved research gaps, add web sources with provenance, create or refine claims, link evidence, challenge conclusions, update confidence with a reason, compare alternatives, and record a draft or final decision. Every mutation becomes immediately visible in the same interface the human is using.
This makes WebMCP the collaboration layer of the product rather than an optional wrapper around a chatbot.
A better research experience
A normal research session often ends up fragmented across tabs, notes, and chat threads. Later it becomes difficult to answer basic questions:
- Which source supported this claim?
- What contradicted it?
- Why did confidence change?
- What remains unknown?
- What evidence would change the decision?
DeepTrail keeps those answers in one durable, inspectable workspace.
Critical-thinking features
Attack this conclusion turns adversarial research into a product workflow. DeepTrail captures a confidence baseline and asks the agent to search specifically for evidence that could falsify or materially qualify the current conclusion.
What would change my mind? lets the investigation record explicit reversal criteria before more evidence arrives.
Research Debt is a deterministic score that surfaces unsupported claims, thin provenance, unresolved questions, and missing counterevidence.
The evidence graph makes relationships between sources and claims visible, while confidence history records why confidence changed rather than silently overwriting a number.
What humans and agents can do together
- A human creates an investigation and defines the question or decision.
- The browser agent reads the structured workspace through WebMCP.
- The agent researches the web and writes credible sources, claims, evidence links, gaps, counterarguments, comparisons, and decisions back into DeepTrail.
- The human sees the same state update immediately and can challenge, edit, or redirect the investigation.
- Both sides can keep unresolved uncertainty visible instead of forcing a premature answer.
WebMCP implementation
DeepTrail uses the imperative WebMCP API with document.modelContext.registerTool() and a state-aware toolset. Tool availability changes with the workspace state so the agent receives only relevant actions.
The implementation includes strict execution-time Zod validation, bounded inputs and outputs, HTTP(S)-only source URLs, same-origin tool exposure, WebMCP annotations, AbortController lifecycle cleanup, compact tool results, stable entity IDs, actor attribution, and local persistence in IndexedDB.
External research text is treated as untrusted content. DeepTrail includes advisory prompt-injection risk checks and does not grant external content authority over tool behavior.
Reliability and privacy
DeepTrail requires no paid LLM API, search API, database, or backend. Research state is stored locally in IndexedDB. Users can export and restore validated JSON backups, and workspace integrity checks reject broken references and malformed imports.
Judge quick start
Open the live /judge route. It contains an evidence-backed seeded investigation, production-readiness diagnostics, and the exact agent prompt for the demo. Load the seeded workspace, send the prompt to a WebMCP-capable browser agent, and watch the agent research a real gap, add evidence, challenge the current draft conclusion, and update the same visible workspace.
The seeded investigation is intentionally imperfect: it includes supporting evidence, a qualification/counterargument, confidence history, explicit falsification criteria, and unresolved production-verification gaps rather than presenting a pre-baked perfect answer.
Built for the WebMCP Challenge
DeepTrail was built during the challenge as a WebMCP-native product focused on human-agent collaboration, inspectability, adversarial thinking, and evidence-backed decisions. The goal is not more generated prose; it is a durable research trail showing what supports a conclusion, what weakens it, and what remains unknown.
Built With
- indexeddb
- netlify
- next.js
- react
- react-flow
- typescript
- webmcp
- zod