W Web Task Agent Local-first deep research

LOCAL-FIRST RESEARCH RECEIPTS

Turn a messy web question into a decision package you can inspect.

Web Task Agent keeps the chain from a recommendation back to its sources: durable jobs, snapshots, evidence clusters, contradictions, resumability, and a handoff package a teammate can challenge instead of a loose pile of notes.

Open a decision receipt Read the docs
Source-linked evidence Contradictions visible Stage-level resume Local by default

PRODUCT PROOF

Open a decision receipt before you configure a browser or model.

It is a real, versioned fixture with a decision, explicit invalidation conditions, and source cards. It needs no key, browser session, or live request to inspect.

Open the live decision receipt →
Preview of the Web Task Agent local research receipt and its decision-package workflow
243 Executable workflows Three core workflows plus 240 catalog entries, all with a stable output contract.
5 Review-gated decision packs Plan idea validation, launches, churn, articles, or integrations before work begins.
8 Deterministic demos Inspect launch, competitor, GitHub, technical, and app-research package shapes without credentials.
1 Auditable local state SQLite stores jobs, sources, documents, extractions, graph links, and artifacts.

What It Actually Does

Research, storage, recovery, and operator control in one local tool.

Long-running research jobs

Plan a job, search multiple queries, fetch result pages in batches, extract evidence, synthesize findings, and write a report.

Durable evidence layer

Store sources, documents, snapshots, extractions, evidence clusters, contradictions, and graph links in SQLite for reuse across runs.

Recovery instead of restart

Leases, heartbeats, cache checkpoints, queue recovery, and stage-level resume keep work moving after interruptions.

Workflow packages, not raw scraps

Topic-based folders include reports, briefs, plan files, raw research snapshots, pipeline manifests, and prompt traces.

Queue, worker, and job controls

Queue long jobs, run a local worker, pause or cancel in flight, resume later, rerun from stored configuration, and inspect live logs.

Local dashboard and API

Serve a management UI plus JSON endpoints for jobs, queue state, recoverable runs, detail views, controls, and SSE event streams.

Execution Model

Each run follows a durable research pipeline.

1

Plan

Generate a bounded job plan and research query set.

2

Search

Use a search adapter to gather result candidates per query.

3

Fetch

Open result pages in batches and checkpoint between fetch slices.

4

Extract

Persist readable pages, structured evidence, and raw snapshots.

5

Synthesize

Build summaries from persisted evidence, clusters, and contradictions.

6

Package

Write the report, drafts, workflow brief, and runtime manifests.

Persistence
  • Jobs, steps, leases, controls, queue state
  • Sources, canonical URLs, documents, extractions
  • Evidence clusters, contradictions, graph links
  • Artifact metadata for reports and handoff files
Recovery
  • Stale lease recovery for long jobs
  • Stage-aware resume for search, fetch, and extract
  • Pause, cancel, resume, rerun, retry controls
  • Failure-mode coverage for interrupted checkpoints

Real Workflows

Start with a decision, then pick the smallest useful workflow.

No-key proof

Inspect a research receipt

Open the receipt first, then inspect the brief, report, evidence and manifest contract before you configure an LLM or browser.

npm run start -- demo export \
  workflow-quality-audit
Open the live receipt
Catalog discovery

Choose from 240 focused cases

Find a distinct voice-of-customer, pricing, competitor, retention, or market-entry workflow by decision and domain.

npm run start -- workflow list \
  --category "Pricing and Packaging"
Browse workflow docs
Human review first

Plan a decision pack

Generate a 2–4 step plan without silently opening sites or spending model calls in a chain.

npm run start -- pack plan validate-an-idea \
  --topic "offline document signing"
See more scenarios

Package Shape

A workflow run produces an operator-facing package, not just one markdown file.

report.md
handoff/
  README.md
  package-manifest.json
  research-summary.md
  workflow-brief.md
drafts/
  post-draft.md
  comments-draft.md
plan/
  plan.json
raw/
  research/
runtime/
  llm-prompt-traces.json
  pipeline-manifest.json

Handoff

The brief and README give the short operator view first, before the full report.

Raw snapshots

Raw research JSON and persisted evidence let you audit or extend a run later.

Runtime traces

Prompt/version traces and pipeline manifests make local debugging much easier.

Operator Surface

Work from the terminal or the local dashboard.

CLI
npm run start -- workflow enqueue android-opportunity \
  --topic "expense tracking for local merchants"
npm run start -- worker run --once
npm run start -- queue list
npm run start -- job logs <job-id> --limit 100
Management API
GET  /
GET  /api/health
GET  /api/jobs
GET  /api/jobs/:id
GET  /api/jobs/:id/events
GET  /api/jobs/:id/events/stream
POST /api/jobs/:id/control
GET  /api/queue
POST /api/queue/:id/control
GET  /api/recoverable

Documentation

Inspect the output contract, then choose a workflow or contribute one.

Documentation Examples Charter Roadmap