Web Search Neo
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Use web_info for discovery and observation. Start with topic=capabilities when the compact contract is not already known. Use web_action for one or many ordered mutations. Read action_schema before an unfamiliar action or topic; after validation failure fix the call from that schema, never guess aliases. execute_js takes script (an async function body: return a value, or send a one-line expression), not code; fill takes fields={CSS_selector: value}. Give each task/agent a unique session_id and agent_label on open; reuse only that task's session. A new session opens isolated (profile_mode='current', the user's own Chrome, only when asked). Never close or take over another agent's tab. After a zero-match click inspect fresh page_elements/find and frame context before retrying; React portals alone do not explain missing text. In step render mode an input action applies all mixed keyboard and pointer changes before advancing exactly one frame. node='' on web_action/web_info runs the call on another computer (web_info topic nodes lists them).
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| web_infoB | Read the contract, the page, the console, the network, or search/browser state. Called with no arguments it returns the whole contract, including recipes and
common mistakes, so no external skill file is needed. Every result (dict
payloads) also carries the current local date/time and UTC-offset region
under the top-level |
| web_actionA | Execute 1-32 ordered objects in actions, each with required canonical action from web_info(topic='actions'); other keys are its parameters (url, session_id, ...). type/name/tool/command/op/operation/method aliases are accepted; prefer action. Read web_info action_schema before unfamiliar actions or validation errors. fill uses fields={CSS_selector: value}. Use a unique session_id per task/agent; never close another agent's tabs. summary='min' (or "summary": "min" in one action) returns short results. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 2 tools
The split is essentially execute vs. read/inspect: web_action performs ordered browser/search actions, while web_info reads contracts, page/console/network state, and browser/search status. The two responsibilities are mostly distinct, though web_info's broad 'read the page/search state' scope could overlap with results returned by web_action.
Both tools use a consistent web_<noun> convention: web_action and web_info. The convention is not verb_noun, but it is predictable and uniform across the set.
Two tools is below the typical 3-15 range and feels thin, even though web_action is an intentional multiplexer handling many canonical actions. The design avoids tool explosion, but pushes discovery and complexity into arguments and the contract-reading flow.
A generic action executor plus a contract/state reader can plausibly cover browser/search lifecycle operations, including session use and inspection. However the actual action surface is hidden behind runtime discovery via web_info, so completeness is less self-evident from the exposed tools.