Tamash Help & User Guide

1. What is Tamash?

TAMASH is a Chrome extension that works with WebMCP tools — actions that a web application registers via Chrome's document.modelContext API. When a page exposes tools like these:

add_item_to_cart   → add a product to the shopping cart
apply_coupon       → apply a discount coupon code
place_order        → finalise the order
user_register      → create a new account

TAMASH lets you:

• Inspect those tools and their schemas in the Tool Tester
• Call each tool manually with custom parameters and see the result
• Run an AI agent that drives the page toward a plain-English goal using those tools
• Design flows — saved sequences of tool calls that replay without any AI

---

2. Prerequisites & Chrome Flags

Chrome version

TAMASH requires Chrome 146 or later. Check your version at chrome://settings/help. Update Chrome if needed before proceeding.

Required flag — Experimental Web Platform features

This flag activates document.modelContext, the core Chrome API that WebMCP tools depend on. Without it, TAMASH finds no tools on any page.

Recommended flag — WebMCP for testing

This flag exposes the navigator.modelContext path alongside document.modelContext and enables wider tool discovery, including apps that use the earlier API shape.

Optional flag — WebMCP support in DevTools

Adds a WebMCP panel to Chrome DevTools so you can inspect registered tools directly in the browser without TAMASH.

---

3. Installation

Chrome Web Store (recommended)

1. Visit the Chrome Web Store listing
2. Click Add to Chrome
3. The TAMASH icon appears in your toolbar

Developer / Unpacked install

1. Clone the repository and run:

   npm install
   npm run dev

2. Open chrome://extensions in Chrome
3. Enable Developer mode (top-right toggle)
4. Click Load unpacked → select the extension/ folder

---

4. Try it with VegCart

VegCart is a purpose-built WebMCP-enabled demo app — a fresh fruit and vegetable e-commerce store with 16 registered WebMCP tools covering products, cart, checkout, wishlist, user accounts, and order history. It is the fastest way to see TAMASH working end-to-end.

VegCart WebMCP tools at a glance

Test credentials

5-minute quick start with VegCart

1. Enable Chrome flags (see Section 2) and install the TAMASH extension
2. Open VegCart in Chrome
3. Click the TAMASH icon → open the 🔧 Tool Tester tab → click ↺ Refresh Tools → 16 tools appear
4. Click get_product_list → click ▶ Run → see all 24 products returned
5. Click add_item_to_cart → enter itemName = "Tomatoes", quantity = 2 → click ▶ Run
6. Switch to the 🤖 Agent tab and type the goal below:

Login as priya@example.com / test123, add Alphonso Mango ×2 and Spinach ×3 to the cart, place an order to 12 Lotus Colony Bengaluru 560001, then show me the order history

The agent runs automatically — login, clear cart, add items, checkout, verify — all from a single prompt.

---

5. The Side Panel

Click the TAMASH icon in your Chrome toolbar to open the side panel. It stays alongside your browser tab as you navigate.

Tab	Purpose
Settings	Configure AI provider, API key, and Remote MCP servers
🔧 Tool Tester	Manually inspect and call WebMCP tools on the current page
🤖 Agent	AI-powered agentic loop that drives WebMCP tools toward a goal
🔀 Flow	Design, save, and replay reusable sequences of WebMCP tool calls

---

6. Settings Tab

Configure your AI provider, API key, and Remote MCP connections. The agent uses these settings for every run.

AI Provider & Credentials

Provider	Models	Notes
OpenAI	gpt-5.4, gpt-5.4-mini, gpt-5.4-nano	API key from platform.openai.com
Anthropic	claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5	API key from console.anthropic.com
Google	gemini-2.5-pro, gemini-2.5-flash, gemini-3.1-pro	API key from aistudio.google.com
Ollama	qwen3.5, qwen3-coder-next, deepseek-v3.2 and more	Default: https://ollama.com. For local: http://localhost:11434

1. Select a Provider — Base URL fills automatically
2. Select a Model
3. Paste your API Key
4. Click Test API Key — a ✓ toast confirms success
5. Click Save Credentials

Your API key is stored in chrome.storage.local on this device only. It is never synced to other Chrome profiles.

Once saved, the form locks to prevent accidental edits. Click Edit to modify any field.

---

7. Tool Tester Tab

The Tool Tester lets you browse every WebMCP tool the current page has registered, inspect its schema, fill in parameters, and call it — without writing any code or running the agent.

How to use it

1. Navigate to a page that registers WebMCP tools (requires Chrome 146+ with the flag enabled)
2. Click the 🔧 Tool Tester tab
3. Click ↺ Refresh Tools — the list populates with all tools on the active page
4. Click any tool name — the right panel shows its description, parameters (type-annotated, required fields marked *)
5. Fill in values and click ▶ Run
6. The result or error appears directly below

The tool list auto-refreshes when you switch tabs or the page reloads.

Example (VegCart)

16 WebMCP tools:
  add_item_to_cart    Add a product to the shopping cart
  get_cart_contents   View current cart items and totals
  place_order         Checkout with delivery address
  user_login          Login with email and password
  ... 12 more tools

→ Click: add_item_to_cart
  Parameters:
    itemName *  (string)   Tomatoes
    quantity    (number)   2

→ Click ▶ Run
→ Result: { "success": true, "message": "Added 2 × Tomatoes to cart", "cartCount": 2 }

---

8. Agent Tab

The Agent tab runs a fully built-in AI agentic loop. Describe a goal in plain language, press Run, and the agent calls WebMCP tools on the current page — turn by turn — until the goal is complete or 25 turns are exhausted.

No external server is needed. The agent runs entirely inside the extension using your configured LLM.

Running the agent

1. Make sure your provider and API key are saved in the Settings tab
2. Navigate to a page with WebMCP tools
3. Click the 🤖 Agent tab
4. Type your goal in the text area
5. Click ▶ Run or press Ctrl+Enter

The agent workspace shows each step live — thinking indicator, tool call chips with arguments and results, and a final result panel. Press ■ Stop at any time to abort.

Agent runtime panel

The collapsible Agent Runtime bar at the top of the Agent tab shows the current provider, model, active page, page tool count, and Remote MCP status. A green alert means ready to run; orange means credentials are missing or tools are unavailable.

How the agent thinks

1. Load tools — fetches all WebMCP tools from the current page plus any configured Remote MCP servers
2. Plan — sends the goal and tool manifest to the LLM
3. Execute — calls the tools the LLM requests; results feed back into the conversation
4. Repeat — up to 25 turns; stops when the LLM responds without tool calls
5. Report — the LLM's final message becomes the run result

Saving a run as a Flow

When an agent run finishes successfully, a 🔀 Save as Flow button appears at the bottom of the run card. Clicking it opens the Flow tab with all captured tool calls pre-loaded — ready to name and save for future replay without AI.

Native agent tool

Tool	What it does
__agent_getTabInfo	Returns the current tab's URL, title, and tabId

Example goals (VegCart)

Provider-specific notes

• OpenAI — standard function calling. Works with any model that supports tools.
• Anthropic — uses direct browser API with anthropic-dangerous-direct-browser-access: true. Works with claude-opus-4-6 and other Claude models.
• Google Gemini — uses functionDeclarations format. Works with Gemini 2.5 and 3.x models.
• Ollama — defaults to https://ollama.com. Override Base URL to http://localhost:11434 for local Ollama. No API key required for local. Use a tool-capable model like qwen3.5 or qwen3-coder-next.

---

9. Flow Tab — Reusable Flows

The Flow tab lets you build, run, import, and export sequences of WebMCP tool calls that replay on demand — no AI needed at run time.

Flow list view

When you open the 🔀 Flow tab you see your saved flows. From here:

• + New Flow — create a blank flow and open the editor
• ⬇ Import — import one or more .json flow files. Also accepts bundles with multiple flows.
• ⬆ Export All — export all flows as a single bundle JSON
• Checkboxes — select individual flows for bulk Export or Delete

Flow editor

Click any flow (or open a new one) to enter the editor. Key actions:

Action	How
Run all steps	Click ▶ Run
Stop mid-run	Click ■ Stop
Save changes	Click ✓ Save
Add a step	Select a tool, fill params, click + Add Step
Edit a step's params	Click the ✎ button on the step
Reorder steps	Drag the ⠿ handle
Delete a step	Click ✕ on the step (with confirmation)
Export / Delete flow	Toolbar buttons at the top of the editor

Flow JSON format

Flows are plain JSON and can be hand-authored or shared:

Single flow (VegCart example):

{
  "name": "Guest Quick Order – Mumbai",
  "steps": [
    { "toolName": "clear_cart",       "params": {} },
    { "toolName": "add_item_to_cart", "params": { "itemName": "Banana",  "quantity": 2 } },
    { "toolName": "add_item_to_cart", "params": { "itemName": "Onions",  "quantity": 1 } },
    { "toolName": "get_cart_contents","params": {} },
    { "toolName": "place_order",      "params": { "name": "Guest User", "phone": "9999999999",
                                                   "address": "1 Marine Drive", "city": "Mumbai",
                                                   "pincode": "400001" } }
  ]
}

Bundle (multiple flows in one file):

[
  { "name": "Flow A", "steps": [ ... ] },
  { "name": "Flow B", "steps": [ ... ] }
]

Creating a flow from an agent run

After a successful agent run, click 🔀 Save as Flow at the bottom of the run card. The Flow tab opens with all captured steps pre-loaded. Name the flow and save it.

---

10. VegCart Flow Library

All 12 flows below are pre-built for VegCart and ready to import into TAMASH. Download any JSON file, open the 🔀 Flow tab, click ⬇ Import, and select the file. The flow runs immediately against the live VegCart tab — no AI, no prompts.

---

11. Remote MCP Servers

Remote MCP Servers let the built-in agent call tools from Jira and GitHub alongside the page's WebMCP tools in the same conversation.

What this enables

Goal: "Find the open Jira bug I'm looking at, then add the item it describes to the cart"

Agent will:
1. jira__search_issues → find the Jira ticket
2. add_item_to_cart    → call the page's WebMCP tool
3. Report success

Adding a remote MCP server

1. Open Settings → scroll to 🔗 Remote MCP Servers
2. Click + Add Remote MCP
3. Choose a provider:
   • Jira Remote MCP — fixed endpoint https://mcp.atlassian.com/v1/mcp
   • GitHub Remote MCP — fixed endpoint https://api.githubcopilot.com/mcp/
4. Fill in credentials, click Test (must succeed), then Add Remote MCP

Provider auth reference

Provider	Auth type	Badge	Credentials
Jira	Personal token	personal-token	Atlassian email + API token
Jira	Service account	service-account	Bearer API key
GitHub	PAT	pat	GitHub personal access token

Tool naming

Remote tools are prefixed with the server name to avoid collisions with page tools:

jira__create_issue
jira__search_issues
github__list_prs
github__create_comment

Agent status messages

12 tools in manifest (8 WebMCP, 4 remote MCP)
🔗 jira: 6 tools
🔗 github: 6 tools

If a server is unreachable, the agent logs a warning and continues with the remaining tools:

⚠ jira unreachable: HTTP 401: Unauthorized

---

12. How WebMCP Tools Work

WebMCP is Chrome's API for websites to declare their own tools for AI agents. A page registers tools by calling document.modelContext.registerTool().

TAMASH intercepts those registrations at page load time and makes the tools available in the Tool Tester, Agent, and Flow tabs.

How TAMASH discovers tools

Three sources are merged in priority order:

1. Interceptor cache (window.__tamash_tools) — TAMASH's content-interceptor.js runs in the MAIN world at document_start and wraps registerTool so every registration is captured before Chrome's native API.
2. Native API (document.modelContext.getTools()) — fills in any declarative tools Chrome registered at parse time.
3. DOM scanner — reads [toolname] attributes on form elements for the WebMCP declarative HTML pattern.

What a WebMCP tool looks like (page code)

const mc = document.modelContext;
mc.registerTool(
  {
    name: 'add_item_to_cart',
    description: 'Add a product to the shopping cart',
    inputSchema: {
      type: 'object',
      properties: {
        itemName: { type: 'string', description: 'Product name' },
        quantity:  { type: 'number', description: 'How many to add' }
      },
      required: ['itemName']
    }
  },
  async (argsJson) => {
    const { itemName, quantity = 1 } = JSON.parse(argsJson);
    // ... perform the action ...
    return JSON.stringify({ success: true });
  }
);

Declarative HTML tools

Pages can also declare tools without JavaScript using HTML attributes:

<form toolname="search_products"
      tooldescription="Search for products by keyword"
      action="/search" method="get">
  <input name="q" toolparamdescription="Search keywords" required>
</form>

---

13. Real-World Examples

Example 1 — Inspect VegCart tools in Tool Tester

1. Open VegCart in Chrome
2. Click the TAMASH icon → 🔧 Tool Tester → ↺ Refresh Tools
3. Select add_item_to_cart → enter itemName = "Tomatoes", quantity = 2 → click ▶ Run

Result: { "success": true, "message": "Added 2 × Tomatoes to cart", "cartCount": 2 }

Example 2 — Agent runs a full VegCart checkout

Goal entered:

Login as priya@example.com / test123, clear the cart, add Alphonso Mango ×2 and Spinach ×3, place an order to 12 Lotus Colony Bengaluru 560001, and show me the order history

What the agent does:

1. Loads 16 WebMCP tools from VegCart
2. user_login → priya@example.com ✓
3. clear_cart ✓
4. add_item_to_cart Alphonso Mango ×2 ✓
5. add_item_to_cart Spinach ×3 ✓
6. get_cart_contents → verifies total ✓
7. place_order Bengaluru 560001 ✓
8. get_order_history → new order confirmed ✓
9. Final result: "Order placed successfully. Cart contained 5 items. Order appears in history."

Example 3 — Save agent run as a reusable flow

1. Complete Example 2 above — the run finishes with "Done"
2. At the bottom of the run card, click 🔀 Save as Flow (8 steps)
3. Name it "Priya Full Checkout" → click ✓ Save

Next time, open the Flow tab → click Priya Full Checkout → press ▶ Run. The same 8 steps execute without any AI or prompting.

Example 4 — Import Flow 11 (Greens Weekly Pack) and run it

1. Download flow-11-greens-weekly-pack.json
2. Open VegCart → TAMASH → 🔀 Flow tab → ⬇ Import → select the file
3. The flow opens in the editor. Click ▶ Run

The flow logs in as Priya, filters to Greens, adds 5 items, reviews the cart, and places the order — 11 steps, no AI.

Example 5 — Cross-system task: search Jira, then add to VegCart

Prerequisites: Jira Remote MCP configured in Settings.

Goal entered:

Search for open BUGS-project Jira tickets titled "Missing item" and add each item name to the VegCart shopping cart

The agent: calls jira__search_issues → for each issue calls add_item_to_cart → calls get_cart_contents → reports items added.

Example 6 — Build a wishlist flow manually

1. Open VegCart → TAMASH → 🔀 Flow → + New Flow
2. Select filter_products_by_category, set category = "fruits", click + Add Step
3. Select add_to_wishlist, set itemName = "Alphonso Mango", click + Add Step
4. Select add_to_wishlist, set itemName = "Pomegranate", click + Add Step
5. Select get_wishlist, click + Add Step
6. Name the flow "Weekend Fruits Wishlist" → click ✓ Save → press ▶ Run

---

14. Tips & Troubleshooting

Tool Tester shows no tools

• Make sure Experimental Web Platform features is enabled at chrome://flags
• Relaunch Chrome after enabling the flag
• Reload the page after relaunching — TAMASH's interceptor must run before the page's scripts fire
• The page must call document.modelContext.registerTool() — not all pages do
• Using VegCart? Make sure you're on vibetestq.com/testweb/vegcart/ (not a local file) and the page has fully loaded before clicking Refresh Tools

Agent says "No tools found" or "tools unavailable"

Same causes as above. TAMASH works only on pages that expose WebMCP tools. Navigate to a WebMCP-enabled page before running.

Agent is stuck or taking too long

Click ■ Stop to abort. The agent has a 25-turn hard limit — it stops automatically when reached. Restart with a more specific goal. Check the Execution log to see where it got stuck.

My API key test fails

• OpenAI: get a key at platform.openai.com
• Anthropic: get a key at console.anthropic.com
• Google: get a key at aistudio.google.com
• Ollama (local): no API key required — leave the field blank and set Base URL to http://localhost:11434
• Keys are stored per device and not synced. Set the key in each Chrome profile separately.

Remote MCP server shows "unreachable"

• For Jira personal token: confirm you are using the Atlassian API token (not your password) with the correct email
• For GitHub PAT: confirm the token has the required scopes and has not expired
• Wrong token → HTTP 401. Verify credentials and click Test again.

Flow fails on a step

The error in the run log shows which tool failed and why. Open the Tool Tester to call that tool manually and debug the parameters. Use ✎ on the failing step to adjust its values.

VegCart: flow runs but nothing happens on the page

• Make sure the VegCart tab is the active tab when you press Run — TAMASH always operates on the currently active tab
• If VegCart reloaded between runs, click ↺ in the Flow editor to refresh the tool list before re-running
• For login flows, confirm you are using the correct credentials: priya@example.com / test123 or ravi@example.com / demo123

Still stuck?

Contact us at support@vibetestq.com or open an issue at github.com/QtpSudhakarOrg/tamashwm.