aisearch
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| WEB_RENDER | No | Render escalation mode for browser rendering: auto, all, or off. | auto |
| OBSCURA_BIN | No | Path to the Obscura binary. Optional; without it, browser rendering and browser_* tools are disabled. | |
| WEB_API_KEY | No | Bearer token used to authorize HTTP API requests. | |
| WEB_HTTP_PORT | No | Port for the HTTP layer. | 8787 |
| WEB_RENDER_MAX | No | Maximum number of browser renders per query. | 3 |
| WEB_CONCURRENCY | No | Concurrency level used by the search engine. | |
| WEB_SERP_ENABLED | No | Enable or disable SERP integration in the search pipeline. | |
| WEB_SOURCE_LOCAL | No | Configuration for local/source domains used by the engine. | |
| WEB_COVERAGE_THRESHOLD | No | Coverage threshold used by the search engine. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| web.searchB | AiSearch: szukaj w zamkniętym ekosystemie (plan → propose → resolve → falsify → [SERP] → learn). Model zgaduje adresy, HTTP je weryfikuje i naprawia (redirect/Wayback). JS-shelle renderuje przeglądarka (auto), gdy treść < renderMinText. Wynik: ranked + provenance + ślad decyzji. |
| web.briefB | Jedno wywołanie → zwinięty brief: ranking + provenance (repaired/rendered/fromCache) + coverage + uwagi. Do raportów, nie do debugu. |
| web.verify_linksC | Partia kandydatów w JEDNYM wywołaniu: survived / rejected / repaired (dowód sensu auto-korektora). serp domyślnie wyłączony — tu celowo bez wynalazców. |
| web.drift_reportC | Raport driftu dokumentacji + prior źródeł + werdykt w języku naturalnym. Domeny, które zaczęły przekierowywać = docs się przestawiły. |
| web.renderB | Render jednej strony przez przeglądarkę (V8): JS, SPA, treść za interakcją. Zwraca markdown + tytuł. Drogie — używaj świadomie. |
| web.traceC | Pełny ślad etapów w języku naturalnym (to, co dashboard rysuje) + statystyki. Do debugowania „co się zadziało". |
| web.feedbackC | Pętla użyteczności: used/opened/ignored/rejected → linklog → prior. Bez tego system uczy się rozstrzygać linki, a nie pomagać. |
| web.evalC | Ewaluacja jakości (replay offline / live) + opcjonalnie werdykt bramki regresji. |
| web.statsC | Wnętrze ekosystemu: cache, prior, drift, liczba wyszukań w sesji. |
| web.driftC | Detektor driftu: domeny z przyspieszonymi przekieraniami. |
| web.cacheC | Stan lub czyszczenie cache (resolve/content/feeds). |
| web.feedsC | Rejestr kanałów RSS/Atom (warstwa świeżości): list/add/remove/reindex. |
| web.configureC | Rozstrzygnięta konfiguracja + stan renderera. |
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 13 tools
Most tools occupy distinct niches (search, render, verify_links, feeds, cache, configure), but web.drift and web.drift_report clearly overlap, and web.stats vs web.trace both expose internal ecosystem state with only subtle wording differences. web.brief vs web.search is also a thin distinction (collapsed output for reports vs full ranked result). Descriptions help, but an agent can plausibly misselect among these pairs.
All tools share a clean `web.` namespace with lowercase snake/compound names, which is highly predictable. The minor deviation is mixed grammatical style — some are verbs (search, render, verify_links, configure) and others plain nouns (stats, cache, feeds, drift), plus drift vs drift_report is an inconsistent granularity pair.
13 tools sits comfortably in a well-scoped range and each maps to a recognizable stage of the search/resolve/learn pipeline. It is slightly heavy on introspection tools (stats, trace, cache, configure, eval) relative to the core search surface, but nothing feels gratuitous.
The lifecycle is essentially covered end-to-end: search, verify, brief, render, trace, feedback/learning loop, eval, caching, feeds and configuration all have entry points. Minor gaps include no explicit single-URL content extraction distinct from render, and no direct read of a stored linklog/prior record despite feedback writing to it.