Reason→ Act→ Observe↻

Build AI agents

Build, run, and improve agents that deliver real work — with Miraptor, open source.

Get started View on GitHub
Open source & growing
↗ Quick start guide

The agent system

One agent runs a real task end-to-end — and delivers.
Pay analysis · production & tech leads at medical-device MNCs (Shanghai & Suzhou) US$2.614
Research and analyze the annual-pay distribution for production and technical leads at foreign medical-device firms in Shanghai and Suzhou.
List files · D:/node-sandbox/8f7b6553-ed7d-4703… · 0 items
Search and collect the latest compensation surveys and market rates for production leads (production/plant directors) and technical leads (R&D/tech directors) at foreign-invested medical-device firms in Shanghai and Suzhou.
Web search · Shanghai medical-device MNC production director / plant / R&D director annual pay report 2024 2025 · 5 results
Web search · Suzhou medical-device MNC plant / production director R&D Director pay 2024 2025 · 5 results
Web search · medical-device production / R&D director salary guide · Michael Page Hays Mercer · 5 results
Collected pay samples for leaders at foreign medical-device firms in Shanghai and Suzhou — spanning reports from Career International, Michael Page, Randstad and others, with structured percentile data on base salary, bonus and long-term incentives.
Quantitatively analyze and compare the pay distribution across Shanghai and Suzhou — pay composition, percentiles (P25, P50, P75, P90) and inter-city gaps.
Write and generate the full pay-distribution report, save it as a Markdown file and deliver.
Completed the pay-distribution report for production leads (plant/production directors) and technical leads (R&D/tech directors) at foreign medical-device firms in Shanghai and Suzhou — covering percentile data (P25/P50/P75/P90), pay-structure ratios and inter-city gaps. Shanghai pay runs roughly 12%–18% higher than Suzhou.
shanghai_suzhou_medical…6.4 KB
Add Hangzhou to the comparison.
Synthesizing Hangzhou data…
Admin › Agents › Writer Unsaved Back Save
Writer
Model
Tools4
Web Search
Search the live web and pull the most relevant, up-to-date pages.
People Search
Look up people and their public profiles across the web and job platforms.
Company Search
Search company, industry and market data across data providers.
File Access
Read, write and organize files in the sandbox — where the deliverables live.
MCP
Skills3
market-research
A reusable playbook: size the market, map rivals and assess risk.
salary-benchmark
Benchmark pay across surveys and job posts, then reconcile the ranges.
exec-brief
Distill the findings into a one-page, decision-ready executive brief.
Openers
Draft · draftDraft
Owns the final-mile expression and formatting of deliverables. Delegate to it once the analysis is done and the output needs shaping to the target reader and format (summary / report / bullets / table). It takes the synthesized analysis and produces the final copy — introducing no new facts and making no analytical calls.
This description is used by the task tool to help the LLM pick the right subagent — be specific about what this agent can do.
EN中文EditPreview
## Responsibilities You own the final-mile transformation of completed analysis into polished, audience-ready text. Your job: shape, structure, and express — not analyze, verify, or research. If the analysis you receive is incomplete or ambiguous, flag it immediately rather than improvise. ## Method Before writing, identify three things: target audience (technical/executive/general), delivery format (summary/report/bullet list/table), and tone register (formal/neutral/conversational). Every structural and stylistic choice must serve those three parameters. Structure ruthlessly: open with the most important information, group related points, eliminate redundancy.
Evaluation· Writer
Role88
Clarity85
Boundaries82
Coverage72
Output70
Writing80
Safety75
This delta is well-scoped and genuinely focused on its domain — it avoids tool-usage instructions and stays clear of frame territory. The Method section is its strongest asset, with the three-parameter pre-writing check standing out as concrete, executable rules. Highest-leverage fix: replace 'publication-ready' adjectives with word-count limits, item counts, and heading conventions.
Defects 6 · medium · restating the frame
Optimize from evaluation
Optimization guidance, e.g. keep the safety section, make the tone more formal…
#37
Research and analyze the annual-pay distribution for production and technical leads at foreign medical-device firms in Shanghai and Suzhou.
Coordinator
Survey the annual-pay distribution of production and technical leads at foreign medical-device firms in Shanghai and Suzhou.
Done
R1
Researcher1· Shanghai research
Deep-search and collect annual-pay data for leaders at Shanghai medical-device MNCs…
Done
R2
Researcher2· Suzhou research
Search pay samples and market rates for leaders at Suzhou medical-device MNCs…
Done
Done
FC
FactChecker· Fact-check
Verify each source's credibility and recency; flag doubtful entries and drop stale samples…
Done
Running
SY
Synthesizer· Synthesis
Synthesize the three cities' data — pay-distribution patterns, percentile structure and inter-city gaps — into structured conclusions…
Running
When ready
WR
Writer· Write report
Turn the synthesis into a reader-ready final report, shaped as summary / report / table…
Idle
Researcher1· Search pay data for production/tech leads at Shanghai medical-device MNCs
You are a professional compensation researcher. Deep-search and collect annual-pay data for production leads (production director, VP of Manufacturing, Plant Manager, etc.) and technical leads (tech director, CTO, R&D Director, etc.) at foreign medical-device firms in Shanghai. Scope: job platforms (Liepin, BOSS Zhipin, 51job, LinkedIn, etc.); pay surveys (Mercer, Hay Group, Career International, ManpowerGroup, etc.)…
Search job platforms (Liepin/BOSS/51job) for Shanghai medical-device MNC production director/manager pay
Search job platforms for Shanghai medical-device MNC tech director/CTO/R&D manager pay
Search pro pay surveys (Mercer/Hay Group/WTW/Career Intl/ManpowerGroup) for medical-device data
Search industry media/forums for pay references and VP-level exec pay
Compile all evidence with full citations into a structured file
Completed the evidence search for Shanghai medical-device MNC production and technical leads' annual pay, across 5 stages: job platforms (production), job platforms (technical), pro pay surveys…
salary-evidence-shanghai-…19.2 KB
Task delivered — evidence document complete, no further work pending.
16 steps · 33 calls · 886.5K tokens
Pay analysis · production & tech leads at medical-device MNCs (Shanghai & Suzhou)Completed
[email protected] gemini-3.6-flash Approval required Created 2026/7/22 19:26 Updated 2026/7/30 10:46
OverviewCompactionSteps29SubagentsSystem promptTools
Σ
Total
1.5M
Input
1.5M
Output
23.6k
Cache
write 0 · read 849.6k
Reasoning
0
Messages
6
7d 15h · 29 model steps · ~53.1k tok/step avg · 56% cache hit
Context budget▶ Replay
Input vs. the model window — the dashed line marks where compaction triggers; the harness folds old context to avoid overflowing the window.
Step 29/29in 79.3kout 4221.0M
Fixed overhead 38.3k Low-water 52.0k High-water 104.1k Output reserve 8.0k Compactions —
Tokens per step
Cache hit Cache write Input Output

Ship anywhere

One codebase, everywhere the work happens

Same agent behavior across three surfaces and three sandbox providers. The only thing that changes is where data lives and where tools run.

① Surfaces where the user meets the agent
Web app Hono + React 19

PostgreSQL-backed, streaming SSE, resumable mid-run.

Desktop app Electron + SQLite

Local-first, keys encrypted via safeStorage, offline-capable.

Chrome extension MV3 + WXT

BUA drives your already-logged-in browser via CDP.

② Sandboxes where the tools actually run
Node local dev

Zero-setup loop — tools run in-process for fastest iteration.

Daytona managed VM

Persistent Linux workspace per task, stateful across rounds.

E2B firecracker

Fresh micro-VM per task, ~200ms cold start, ephemeral.

Dependency-inverted — swap a sandbox provider without touching agent code. Capability-gated tools hide what a sandbox can't do.

How it works

Autonomous by default, shaped by protocol

The same protocol — from a one-line answer to a multi-file refactor. Clarify and Plan activate only when Assess flags them as needed.

Assess

Triage against P0-P3 policies — direct answer, confirmation required, clarification needed, or plan first.

Clarify

optional

When context, goal, or constraints materially affect the outcome, ask the user before acting.

Plan

optional

Emit an explicit todo list — execution stages only, each with verifiable output and a delivery spec.

Execute

Run stages linearly, parallelize independent tool calls when safe, recover from errors openly.

Deliver

Synthesize outputs, verify against the delivery spec, return a direct response or persisted files.

FAQ

Answers before you fork it

The things people ask after five minutes with the repo.

Can I self-host?

Yes — Zapvol is open source end-to-end. Run the server + web stack on your own infra, or ship the desktop app straight to users. No phone-home, no mandatory cloud account.

Why TypeScript instead of Python?

The agent shares code with the UI — prompts, schemas, tool types, message shapes — and TypeScript lets us ship the same typed contract from the browser down to the sandbox. Python would force a protocol boundary we don't need.

How is this different from Mastra / Vercel AI SDK / LangChain?

Vercel AI SDK is the streaming primitive underneath — we build on it. Mastra is a peer framework; Zapvol goes further by shipping a full Web + Desktop + Extension product, not just a backend framework. Against LangChain, we favor typed, documented internals you can read in an afternoon.

Is the browser extension (BUA) safe?

BUA is an internal-only tool: silent execution by default, no consent prompts, but a full audit log, a global kill switch in the popup, and Chrome's non-removable yellow "debugger attached" bar. A domain blocklist lets ops carve out sensitive sites.

Which models are supported?

Anything Vercel AI SDK supports — Anthropic, OpenAI, Google, Bedrock, Azure, local Ollama. Model choice is per-tier; tier configs are stored in the DB and swappable without a deploy.

How does sandbox isolation work?

Every tool call resolves against a sandbox capability map. The agent never touches your host filesystem or shell directly — it calls a capability, the sandbox provider (Node / Daytona / E2B) executes it, and the result streams back as a typed tool output.

Is this production-ready?

Zapvol is used in production internally for coding, research, and browser automation. The stack — streaming, resumable rounds, 3-tier compaction, per-task SSE replay via Redis — is built for long-running agent work, not demos.

What's the license?

Source-available under a license that permits self-hosting and internal use. See the repo for the exact terms; commercial redistribution needs a separate agreement.

Ready to build?

Get started with Zapvol in minutes. Free and open source.

View on GitHub Read the Docs