AI-native browser runtime for autonomous agents with human-in-the-loop supervision
io.github.unmodeled-tyler/vessel-browser MCP Server
This MCP server provides an AI-native browser runtime intended for autonomous agents, with human-in-the-loop supervision. The project is associated with “mcp-server” functionality and is positioned for AI web browser automation using agent and LLM agent workflows.
🛠️ Key Features
AI-native browser runtime
Autonomous agents with human-in-the-loop supervision
MCP server integration for browser-based agent workflows
🚀 Use Cases
AI browser automation for autonomous agents
Human-supervised agent tasks that require web browsing
LLM-agent-driven interactions with web content
⚡ Developer Benefits
Works with agent and LLM-agent ecosystems via MCP
Supports BYOK (as listed in topics)
Includes platform scope for macOS universal and Linux (as listed in topics)
⚠️ Limitations
Server capabilities beyond the provided description, including tool inventory, are not specified in the available data.
Vessel is a simple, clean web browser with an AI assistant built in. Browse the web normally, then ask Vessel to help research, navigate pages, summarize what you are reading, fill forms, or keep track of longer tasks.
Unlike invisible browser automation, Vessel shows you what the AI is doing in a real browser window. You can follow along, approve important actions, pause or redirect the work, and take over whenever you want.
A familiar browser, with AI help nearby — tabs, bookmarks, reader mode, downloads, and a sidebar assistant for web tasks
Great for longer browsing sessions — durable conversation threads, a cross-source run inbox, named sessions, checkpoints, notes, bookmarks, and page changes make interrupted work recoverable
You stay in control — review what the AI is doing, approve sensitive actions once or for a scoped run/domain, reject with steering, undo recent changes, and switch back to manual browsing at any time
Open source and extensible — built on Chromium, with advanced automation support for MCP clients and agent tools when you need it
Linux is the most mature install target today. macOS release packaging is available from source.
Vessel is in active development. Treat it as early software, especially for sensitive browsing or account access.
Vessel development uses Node.js 22. If you use fnm, run fnm use from the repo root to pick up .node-version.
bash
fnm use
npm install
npm run dev
If you want extra local AI tracing in development, create an optional src/main/telemetry/trace-logger.local.cjs file. Vessel will load it only in local dev builds, and packaged production builds ignore it.
Why Vessel?
Most AI tools can answer questions about the web. Vessel is for actually working on the web: reading pages, following links, comparing information, filling forms, tracking changes, and keeping longer browsing sessions organized.
Vessel Browser supervisor sidebar showing agent workflow controlsVessel Browser command bar and agent-driven page interactionVessel Browser persistent session and checkpoint interfaceVessel Browser integrated chat assistant with browser tools
Vessel keeps that work visible. The AI uses a real browser, so you can see the pages it opens, the choices it makes, and the places where it needs your approval. You are never stuck watching a hidden automation run or guessing what happened after the fact.
For power users and developers, the same foundation also supports external agents, MCP clients, persistent sessions, checkpoints, and browser automation workflows.
Features
AI-assisted browsing — ask Vessel to help with research, page reading, navigation, form work, and multi-step web tasks
Human-visible browser UI — pages render like a normal browser so AI activity stays legible instead of disappearing into a hidden run
Command Bar (Ctrl+L) — a secondary operator surface for harness-driven workflows and future runtime commands
Supervisor Sidebar (Ctrl+Shift+L) — live supervision across ten tabs, including a durable Run Inbox, conversation Threads, approvals, checkpoints, changes, and Research
Chat Assistant — built-in conversational AI in the sidebar Chat tab; supports Anthropic, OpenAI, Ollama, llama.cpp, Mistral, xAI, Google Gemini, OpenRouter, and any OpenAI-compatible endpoint; reads the current page automatically; has full access to the same browser tools as external agents; multi-turn session history; configure provider, model, and API key in Settings
Skills (Premium) — reusable browser skills in the sidebar Skills tab; import, view, run, or delete skills the built-in agent can use for research, shopping, and site-specific workflows
Research Desk (Beta) — a dedicated sidebar Research tab for structured research reports; start with a topic, complete an in-tab briefing, let Vessel draft research objectives, approve the plan, then dispatch browser sub-agents to collect source-backed claims and synthesize a markdown-exportable report. Starting the brief is free; plan approval, sub-agent execution, and report export require Vessel Premium.
Dev Tools Panel (F12) — inspect console output, network requests, and MCP/agent activity in a resizable panel at the bottom of the window; export logs by category and date range as JSON
Browser Basics For Long Runs — pinned tabs stay compact at the front of the tab strip and are protected from accidental close; tab groups can be color-coded and collapsed; audible tabs show audio indicators with mute controls; open additional browser windows with Ctrl+N; print the active page with Ctrl+P or save it directly as PDF with Ctrl+Shift+P
Action Undo / Rollback — restore the browser to the session snapshot captured immediately before the last successful mutating agent action; available from the Supervisor tab and through the undo_last_action tool
Durable Run Inbox — Chat, MCP, scheduled, and Research work share one persisted lifecycle with status filters, parent/child research runs, redacted inputs, outputs, errors, and tab context
Persistent Conversation Threads — group multiple independent chats inside named threads; each chat gets an AI-generated title after its first completed exchange and can be reopened or retitled, while threads can be renamed, archived, or deleted; archived threads follow the configured history period
Scoped Approval Policies — approve once, approve for a run, approve for a domain, or reject and steer; explicit denies take precedence and expired scoped rules are ignored
Agent-Meaningful Bookmarks — bookmarks carry structured context the agent can read and act on: intent (what the page is for), expectedContent (what to expect on the page), keyFields (important form fields), agentHints (arbitrary directives), and a stored pageSchema; humans can create and edit this metadata directly in the Bookmarks tab, and all fields are searchable
Portable Bookmark Export — export browser-compatible Netscape HTML for import into Chrome, Firefox, Safari, Edge, Brave, and other browsers; optionally include Vessel notes/agent metadata, or export a full-fidelity Vessel JSON archive
Page Schema Inference — Vessel automatically infers a typed schema for every page: pageType (article, product, form, search, checkout, login, dashboard), primaryEntity (structured fields for products and articles), formFields (with names, types, labels, selectors), and actionButtons (with inferred intents: submit, addToCart, login, etc.); schema is attached to every content extraction result
Bookmarks for Agents — save pages into folders, attach one-line folder summaries, and search bookmarks over MCP instead of dumping the entire library
Named Session Persistence — save cookies, localStorage, and current tab layout under a reusable name, then reload it after a restart
Annotated Checkpoints — capture and restore short-lived browser recovery points with names and editable notes, so humans and agents can mark why a checkpoint matters before risky flows
Page Highlights — agents can visually highlight text or elements on any page with labeled, color-coded markers that persist across navigation; highlight count and navigation controls appear in the sidebar; cleared explicitly or via tool call
Agent Transcript Dock — floating transcript overlay anchored to the browser chrome; configurable display modes (off, summary, full) set in Settings; shows live agent thinking and status updates without occupying sidebar space
Workflow Flow Tracking — agents can declare a named multi-step workflow at runtime using flow_start; progress is tracked step-by-step with flow_advance and visible in the sidebar throughout execution
Structured Page Visibility Context — extraction can report in-viewport elements, obscured controls, active overlays, and dormant consent/modal UI
Popup Recovery Tools — agents can explicitly dismiss common popups, newsletter gates, and consent walls instead of brute-forcing generic clicks
Form Autofill Profiles — save reusable personal or work profiles in Settings and fill common contact, address, and organization fields on the current page; Vessel matches fields using labels, names, placeholders, and autocomplete hints
Page Diff / "What Changed?" — Vessel remembers the last snapshot of a page and surfaces a Changed badge in the address bar when the title, headings, or main content differ on a later visit; expand it to see a compact summary of what changed since the last snapshot
What Changed Timeline (Premium) — the sidebar Changes tab keeps a per-page history of recent change bursts, showing when each update was detected and a compact summary of what changed
Per-Tab Ad Blocking Controls — tabs default to ad blocking on, but agents can selectively disable and re-enable blocking when a page misbehaves
Domain Policy — allowlist or blocklist domains globally in Settings; agents cannot navigate to blocked domains
Agent Credential Vault (Premium) — encrypted credential storage for agent-driven logins; credentials are filled directly into login forms via a "blind fill" pattern and are never sent to AI providers; user consent dialog before every use; TOTP 2FA support; domain-scoped access; append-only audit log
Screenshot & Visual Analysis (Premium) — take a full-page screenshot and pass the image directly to the AI for visual layout analysis; useful when text extraction fails on heavy or canvas-rendered pages
Obsidian Memory Hooks (Premium) — optional vault path for agent-written markdown notes, page captures, and research breadcrumbs
Runtime Health Checks — startup warnings for MCP port conflicts, unreadable settings, and user-data write failures
Reader Mode — extract article content into a clean, distraction-free view; toggle on and off from the address bar
Focus Mode (Ctrl+Shift+F) — hide all chrome, content fills the screen
Resizable Panels — drag the sidebar edge to resize; width persists across sessions
Minimal Dark Theme — warm dark grays, restrained accent color, and no pure black/white
Reliability, History, and Approvals
The Supervisor sidebar now keeps longer-running work understandable across Chat, external MCP agents, scheduled jobs, and Research:
Open Runs to filter work that needs attention, is active, failed, completed, or comes from any source. Each run has a chronological timeline with status changes, approvals, outputs, errors, and relevant tab context; records can be exported as Markdown or JSON or deleted from Vessel.
Open Threads to organize multiple chats under a named topic. Chats can be reopened or retitled independently, while their parent thread can be renamed, archived, or deleted. New chats receive a generated title after their first completed exchange.
Approval cards show the action, redacted arguments, domain, and whether undo is available. You can approve once, approve matching work for the current run or domain, reject, or reject with written steering. Persisted scoped rules remain visible in Supervisor and can be removed there.
If Vessel closes while a run is active or waiting for approval, that run is marked interrupted on the next launch instead of being left misleadingly active. Vessel does not claim to resume or retry the unfinished execution automatically.
Terminal run history and archived threads follow historyRetentionDays (90 days by default; supported values are 7, 30, 90, 180, 365, or null for no automatic expiry). Active runs and unarchived threads are preserved.
Conversation history is encrypted through Electron's safeStorage API when encryption is available. If secure storage is unavailable, Vessel keeps conversations in memory for that session instead of writing plaintext chat history. Run records redact sensitive argument fields and limit stored output size before persistence or export.
Positioning
Most browsers treat automation as secondary and assume a human is the primary actor. Vessel is the opposite: it is the browser for the agent, with a visible interface that keeps the human in the loop.
That means the product should optimize for:
persistent browser state across tasks and sessions
clear visibility into what the agent is doing right now
lightweight human intervention instead of constant manual driving
a browser runtime that can serve long-lived agent systems such as Hermes Agent or OpenClaw-style harnesses
Stack
Layer
Technology
Engine
Chromium (Electron 40)
UI Framework
SolidJS
Language
TypeScript
Build
electron-vite + Vite
AI Control
External agent harnesses (Hermes Agent, OpenClaw, MCP clients) + built-in chat (Anthropic, OpenAI, Ollama, llama.cpp, and any OAI-compatible endpoint)
Content Extraction
@mozilla/readability
Architecture
code
Main Process Renderer (SolidJS)
├── TabManager (WebContentsView[]) ├── TabBar, AddressBar
├── AgentRuntime (session + supervision) ├── CommandBar (secondary surface)
├── Run, conversation, and policy managers ├── AI Sidebar (Supervisor/Runs/Threads/Chat/...)
├── MCP server for external agents ├── DevTools Panel (Console/Network/Activity)
├── AI providers (Anthropic + OAI-compatible) ├── Agent Transcript Dock
├── Bookmarks, checkpoints, and page history └── Signal stores (runtime/runs/conversations/policies/...)
└── IPC handlers ◄──contextBridge──► Preload API
Each browser tab is a separate WebContentsView managed by the main process. The browser chrome (SolidJS) runs in its own view layered on top. All communication between renderer and main goes through typed IPC channels via contextBridge.
The sidebar Skills tab renders skill forms entirely in the renderer and passes the rendered prompt to the built-in agent via the same query() path used by the Chat tab — no additional IPC surface is needed. The Changes tab reads the current page's diff timeline through IPC and unlocks persisted history for Premium users. The Research tab has a dedicated IPC surface for its state machine: briefing is available to everyone, while objective approval, parallel browser sub-agents, report synthesis, and markdown export are Premium-gated during the beta.
Getting Started
The installer:
clones or updates Vessel into ~/.local/share/vessel-browser
installs dependencies and builds the app
creates a vessel-browser launcher in ~/.local/bin
creates a vessel-browser-launch helper in ~/.local/bin
creates a vessel-browser-update helper in ~/.local/bin
creates a vessel-browser-status helper in ~/.local/bin
creates a desktop entry for Linux app launchers
writes ~/.config/vessel/vessel-settings.json with MCP port 3100
writes ~/.config/vessel/mcp-stdio-snippet.json
writes ~/.config/vessel/mcp-http-snippet.json
installs a vessel-browser-mcp helper that can run as a stdio-to-HTTP proxy (--stdio) or print config snippets
prints the exact recommended stdio MCP snippet to paste into your harness config
The packaged AppImage path:
does not require a local Node/Electron toolchain
uses the packaged Vessel app icon and metadata
is the recommended path for early adopters who just want to run Vessel
Windows packaged releases:
use the Vessel-<version>-x64-setup.exe NSIS installer
can be installed over an existing Vessel install when upgrading
preserve Vessel app data during the normal upgrade path
You do not need to uninstall Vessel before installing a newer Windows release. Uninstall first only if you are recovering from a broken install or intentionally removing local Vessel data.
After install:
bash
vessel-browser
bash
# Use the pinned Node.js 22 runtime
fnm use
# Install dependencies
npm install
# If Electron download fails, use a mirror:
ELECTRON_MIRROR="https://npmmirror.com/mirrors/electron/" npm install
# Development (with HMR)
npm run dev
# Production build
npm run build
# Smoke-test the MVP release path
npm run smoke:test# Package an unpacked Linux app
npm run dist:dir# Package a Linux AppImage
npm run dist
# Package an unpacked universal macOS app bundle (run on macOS)
npm run dist:mac:dir# Package universal macOS DMG + ZIP artifacts (run on macOS)
npm run dist:mac
# Verify a universal macOS build contains Intel and Apple Silicon slices
npm run verify:mac:universal
# Package signed macOS DMG + ZIP artifacts (run on macOS with signing set up)
npm run dist:mac:signed
Notes:
npm run dev still launches the stock Electron binary, so Linux may continue showing the default Electron gear icon in development
packaged builds created with npm run dist / npm run dist:dir use the Vessel app icon
npm run build:icon:mac regenerates resources/vessel-icon.icns from resources/vessel-icon.png for macOS packaging
npm run dist:mac and npm run dist:mac:dir build universal macOS artifacts by default, so the same app runs natively on Intel and Apple Silicon Macs
npm run dist:mac:x64 and npm run dist:mac:arm64 are available when you need smaller architecture-specific test artifacts
npm run verify:mac:universal uses lipo to confirm the packaged app executable and Electron framework include both x86_64 and arm64 slices
npm run dist:mac, npm run dist:mac:dir, npm run dist:mac:x64, and npm run dist:mac:arm64 intentionally disable auto-signing so local packaging works on any Mac without keychain setup
npm run dist:mac:signed and npm run dist:mac:dir:signed build universal artifacts and use normal electron-builder signing discovery; if your login keychain has duplicate Apple certs, clean those up or use a dedicated keychain before running the signed path
signed builds are still not notarized by this repo out of the box, so Gatekeeper warnings remain until notarization is added for release publishing
the tracked smoke test runs typecheck, build, the MCP stdio proxy regression check, and the Electron navigation regression harness
for headless CI, run the smoke test under xvfb-run -a npm run smoke:test
run npm run test:navigation-regression for the standalone navigation suite; it builds fresh test/preload bundles, opens focused test windows, and uses a temporary profile and loopback server
the navigation runner requires a completed-suite report as well as a zero exit, times out after three minutes, and removes its temporary files on normal completion or failure
Setting up Vessel for Hermes Agent or OpenClaw
Vessel is designed to act as the browser runtime that your external agent harness drives.
Launch Vessel
Open Settings (Ctrl+,) to confirm MCP status, copy the endpoint, or change the MCP port
Optional: set an Obsidian vault path, create autofill profiles, or adjust session preferences
Start Hermes Agent or OpenClaw and point it at Vessel — the easiest way is vessel-browser-mcp --stdio as the MCP command (auth is resolved automatically), or connect directly to http://127.0.0.1:<mcpPort>/mcp with the bearer token from ~/.config/vessel/mcp-auth.json
Use the Supervisor panel in Vessel's sidebar to pause the agent, change approval mode, review pending approvals, checkpoint, undo the last mutating action, or restore the browser session while the harness runs
Open Runs to inspect or export activity from Chat, MCP, scheduled jobs, and Research; use Threads to reopen and organize durable conversations
Use the Bookmarks panel to organize saved pages into folders, edit agent-facing bookmark metadata, export bookmarks for other browsers, and expose saved pages back to the agent over MCP
Notes:
Vessel exposes browser control to external agents through its local MCP server
The MCP endpoint supports both legacy clients and the modern 2026-07-28 protocol revision, including server/discover
The default MCP port is 3100
Hermes Agent and OpenClaw should treat Vessel as the persistent, human-visible browser rather than launching their own separate browser session
Vessel supports a built-in Chat tab with configurable AI provider; open Settings (Ctrl+,) and enable Chat Assistant to set a provider and model
The sidebar Research tab is marked Beta; use it to turn a broad research topic into a brief, approve a multi-thread plan, run browser sub-agents, and export the final report as markdown. Briefing is free, while full execution and export require Premium.
llama.cpp (Local) is a first-class chat provider in Settings and targets http://localhost:8080/v1 by default; Vessel auto-fetches the active model from llama-server
For llama-server, use --ctx-size 16384 minimum and 32768 recommended for reliable Vessel agent loops; lower values often fail once prompt, tool schema, and tool history accumulate
Approval policy is controlled live from the sidebar Supervisor panel rather than a separate global settings screen
Settings now show MCP runtime status, active endpoint, startup warnings, and allow changing the MCP port with an immediate server restart
Settings also include reusable Form Autofill profiles for one-click filling of common contact and address forms on the active page
The address bar can also show a Changed badge when Vessel detects that a previously visited page has meaningfully changed since the last saved snapshot
Premium users can open the sidebar Changes tab for the full What Changed timeline for the active page
The Bookmarks tab can export browser-compatible HTML, HTML with Vessel notes, or a full Vessel JSON archive with agent metadata intact
Agents can selectively disable ad blocking for a problematic tab, reload, retry the flow, and turn blocking back on later
Agents can persist authenticated state with named sessions, for example github-logged-in, and reload that state in later runs
The intended control plane is an external harness driving Vessel through MCP
If you set an Obsidian vault path in Settings, harnesses can write markdown notes directly into that vault via Vessel memory MCP tools
Using llama.cpp as the built-in chat provider
Vessel can talk directly to a local llama-server through its OpenAI-compatible API.
Click refresh if needed; Vessel will auto-detect the active model from http://localhost:8080/v1
Notes:
--ctx-size 16384 is the minimum practical setting for Vessel agent loops
--ctx-size 32768 is the recommended default for longer browsing sessions
Vessel will warn in Settings if it detects a llama-server context size below the recommended floor, or if it cannot detect the ctx size from the running server
Initial memory tools:
vessel_memory_note_create
vessel_memory_append
vessel_memory_list
vessel_memory_search
vessel_memory_page_capture
vessel_memory_link_bookmark
Bookmark and folder tools exposed today include:
vessel_bookmark_list
vessel_bookmark_search
vessel_bookmark_open
vessel_bookmark_save
vessel_bookmark_remove
vessel_create_folder
vessel_folder_rename
vessel_folder_remove
Page interaction and recovery tools exposed today include:
vessel_extract_content
vessel_read_page
vessel_scroll
vessel_dismiss_popup
vessel_set_ad_blocking
vessel_wait_for
vessel_screenshot (Premium) — capture the full page as an image for visual AI analysis
undo_last_action — restore the browser to the snapshot captured before the last successful mutating agent action
Page highlight tools:
vessel_highlight — visually mark text or an element on the page with a labeled, color-coded overlay; persists until cleared
vessel_clear_highlights — remove all highlights from the current page
Workflow tracking tools:
vessel_flow_start — begin a named multi-step workflow and declare its steps upfront; progress appears in the sidebar throughout execution
vessel_flow_advance — mark the current step complete and advance to the next
vessel_flow_status — check current workflow progress
vessel_flow_end — clear the active workflow tracker
Data extraction tools (Premium):
vessel_extract_table — extract a page table as structured JSON rows with column headers
Named session tools exposed today include:
vessel_save_session
vessel_load_session
vessel_list_sessions
vessel_delete_session
Session files are sensitive because they may contain login cookies and tokens. Vessel stores them under the app user-data directory with restrictive file permissions.
Agent Credential Vault tools (Premium):
vessel_vault_status — check whether stored credentials exist for a domain (returns labels/usernames, never passwords)
vessel_vault_login — fill a login form using stored credentials (blind fill — credentials go directly into the page, never into the AI conversation)
vessel_vault_totp — generate and fill a TOTP 2FA code from a stored secret
Session performance tools (Premium):
vessel_metrics — show per-tool call counts, average durations, error rates, and total session stats
Vault security model:
Credentials are encrypted at rest using AES-256-GCM with a key protected by the OS keychain (Electron safeStorage)
Credential values are never sent to AI providers — they flow only through the main process to the content script
Every credential use triggers a user consent dialog ("Allow Once" / "Allow for Session" / "Deny")
All credential access is recorded in an append-only audit log
Credentials are domain-scoped — they can only be used on matching domains
Users manage credentials in Settings > Agent Credential Vault
Notable extraction modes include:
visible_only — only currently visible, in-viewport, unobstructed interactive elements plus active overlays
results_only — likely primary search/result links only
full / summary / interactives_only / forms_only / text_only
The extraction output can distinguish:
active blocking overlays
dormant consent/modal UI present in the DOM but not active for the current session or region
The stdio proxy reads the bearer token from ~/.config/vessel/mcp-auth.json at connection time and preserves both legacy and 2026-07-28 protocol negotiation, so no manual token management is needed.
Vessel must already be running when your MCP client connects, and ~/.config/vessel/mcp-auth.json must exist from install or first launch.
Generic HTTP MCP config (requires copying the token manually):
json
{"mcpServers":{"vessel":{"type":"http","url":"http://127.0.0.1:3100/mcp","headers":{"Authorization":"Bearer <token from ~/.config/vessel/mcp-auth.json>"}}}}
Hermes Agent config.yaml MCP config:
yaml
mcp_servers:vessel:url:"http://127.0.0.1:3100/mcp"headers:Authorization:"Bearer <token from ~/.config/vessel/mcp-auth.json>"timeout:180connect_timeout:30
Configuration
The installer writes three snippets to:
~/.config/vessel/mcp-stdio-snippet.json
~/.config/vessel/mcp-http-snippet.json
~/.config/vessel/mcp-hermes-snippet.yaml
It also installs a helper command:
bash
vessel-browser-mcp
Helper examples:
bash
# Run as stdio-to-HTTP proxy (for MCP client integration)
vessel-browser-mcp --stdio
# Recommended stdio MCP snippet
vessel-browser-mcp
# Generic JSON snippet with Authorization header
vessel-browser-mcp --format json
# Hermes-ready YAML snippet with Authorization header
vessel-browser-mcp --format hermes
# Raw MCP endpoint URL
vessel-browser-mcp --format url
# Raw MCP bearer token
vessel-browser-mcp --format token
Source install update helpers:
bash
# Check whether a source-install update is available
vessel-browser-update --check
# Fetch, rebuild, and update the local source install
vessel-browser-update
Status helper:
bash
# Human-readable local install + MCP status
vessel-browser-status
# Machine-readable status for harnesses
vessel-browser-status --json
Smart launch helper:
bash
# Launch Vessel using the best available local install
vessel-browser-launch
# Show the chosen launch path without starting anything
vessel-browser-launch --dry-run
vessel-browser-launch prefers a healthy source install and falls back to the newest local AppImage when the source install is likely blocked by Electron sandbox permissions.
