Global Settings
Global Settings is the personal configuration page for the current user. Settings configured here apply across all projects. It is accessible from the gear icon in the top navigation bar (far right).
The page is divided into six sections: LLM Providers, Agent Skills, Chat Skills, Tradecraft, API Keys & Tunneling, and System.
Configure the AI model providers that power the agent. All providers added here become available in the model selector of every project's settings.
Each provider card shows its icon, name, type, and — for OpenAI-Compatible entries — the model identifier. You can edit, delete, or test each provider from its card.
Click Add Provider to register a new provider. Choose the type (OpenAI, Anthropic, Google Gemini, DeepSeek, GLM, Kimi, Qwen, xAI, Mistral, OpenRouter, AWS Bedrock, or OpenAI-Compatible), enter your credentials, and test the connection before saving. Each form includes a "Get API key →" link to the provider's official console.
For full details on supported providers, model discovery, and setup guides, see AI Model Providers.
Store API keys for external OSINT and reconnaissance services. These keys are saved per-user in the database and are used by both the AI agent's tools and the recon pipeline at runtime.
| Field | Used by | What it enables |
|---|---|---|
| Tavily API Key | AI Agent |
web_search tool — CVE research, exploit lookups, and general web queries |
| Shodan API Key | AI Agent, Recon Pipeline, Uncover |
shodan tool -- internet-wide OSINT (host info, reverse DNS, domain DNS, passive CVEs) |
| SerpAPI Key | AI Agent |
google_dork tool — Google dorking OSINT (site:, inurl:, filetype:). Free tier: 250 searches/month |
| PDCP API Key | AI Agent |
cve_intel tool — ProjectDiscovery vulnx CVE database (NVD + CISA KEV + EPSS + PoCs + Nuclei templates). Optional: works anonymously at 10 req/min; key lifts the rate limit. Free signup at cloud.projectdiscovery.io. The key is never written to docker-compose.yml, env vars, or build args — it lives only in your user settings and is silently injected per call. |
| NVD API Key | Recon Pipeline | NIST NVD API key — increases CVE lookup rate limit from 5 to 120 requests/30s |
| Vulners API Key | Recon Pipeline | Vulners CVE database — alternative to NVD for vulnerability lookups with richer exploit data |
| URLScan API Key | Recon Pipeline | URLScan.io OSINT enrichment — higher rate limits and access to private scans |
| GitHub Access Token | GitHub Secret Hunt, TruffleHog | Personal Access Token (ghp_...) for GitHub API access — used by both the GitHub Secret Hunt and TruffleHog secret scanning modules. Requires repo scope for private repo access (public repos work without it) and read:org for organization scanning. For maximum security, use a fine-grained token scoped to specific repositories. |
| Censys API Token | Recon Pipeline, Uncover | Censys Platform personal access token -- IP host data, open ports, TLS certs, ASN via Platform API v3. Free tier available at search.censys.io |
| Censys Organization ID | Recon Pipeline, Uncover | Paired with Censys API Token. Found on your Censys account page |
| FOFA API Key | Recon Pipeline, Uncover | FOFA internet asset search enrichment -- IP:port pairs, HTTP titles, server headers, geolocation. Supports legacy (email:key) and modern (key-only) auth formats. Supports key rotation |
| OTX API Key | Recon Pipeline | AlienVault Open Threat Exchange — threat reputation, malware families, passive DNS, MITRE ATT&CK. Optional: OTX enrichment is enabled by default and works anonymously (1,000 req/hr). Adding a key raises the limit to 10,000 req/hr. Supports key rotation |
| Netlas API Key | Recon Pipeline, Uncover | Netlas internet intelligence -- ports, HTTP response data, geolocation, TLS certs, DNS records, WHOIS. Supports key rotation |
| VirusTotal API Key | Recon Pipeline | VirusTotal reputation for domains and IPs — AV verdicts, reputation score, categories, JARM fingerprint. Free tier: 4 requests/minute (configurable). Supports key rotation |
| ZoomEye API Key | Recon Pipeline, Uncover | ZoomEye host search -- ports, service banners, device/OS fingerprints, geolocation, ASN, SSL info. Supports key rotation |
| CriminalIP API Key | Recon Pipeline, Uncover | Criminal IP threat intelligence -- risk score, threat tags (VPN/Tor/proxy/C2/scanner), geolocation, abuse history. Supports key rotation |
These keys are used exclusively by the Uncover target expansion module (GROUP 2b). They are optional -- uncover also reuses any Shodan, Censys, FOFA, ZoomEye, Netlas, or CriminalIP key configured above.
| Field | Used by | What it enables |
|---|---|---|
| Quake API Key | Uncover | 360 Quake cyberspace search -- asset discovery by service, certificate, and banner. Supports key rotation |
| Hunter API Key | Uncover | Qianxin Hunter cyberspace search -- Chinese threat intelligence platform. Supports key rotation |
| PublicWWW API Key | Uncover | Source code search engine -- find websites using specific technologies, scripts, or snippets. Supports key rotation |
| HunterHow API Key | Uncover | hunter.how internet search -- asset discovery and reconnaissance. Supports key rotation |
| Google Custom Search API Key | Uncover | Google Custom Search JSON API (different from SerpAPI) |
| Google Custom Search CX | Uncover | Programmable Search Engine ID -- paired with Google API Key above |
| Onyphe API Key | Uncover | Cyber defense search engine for exposed assets, threat detection, and attack surface management. Supports key rotation |
| Driftnet API Key | Uncover | Fast internet-wide port and service discovery. Supports key rotation |
Each field is a secret input with an eye icon to toggle visibility. Signup links are provided next to each field to help you obtain a key. After entering or updating a key, click Save Settings to persist the change.
Note: All API keys are stored exclusively in the database via this page (user-scoped). They are not read from environment variables or project settings.
Each API key field has a Key Rotation button on the right side. This lets you configure multiple keys per tool for automatic round-robin rotation to avoid rate limits.
Click the button to open the rotation modal:
- Extra API Keys — paste additional keys, one per line. These keys plus the main key form the rotation pool. All keys are treated equally.
- Rotate Every N Calls — after this many API calls, the system switches to the next key in the pool (default: 10).
Once configured, a badge below the field shows the total key count and rotation interval (e.g., "3 keys total, rotate every 10 calls").
Tip: Key rotation is especially useful for NVD (rate-limited to 5 requests/30s without a key), Shodan (paid plans have per-key quotas), the 7 OSINT threat intelligence tools (Censys, FOFA, OTX, Netlas, VirusTotal, ZoomEye, CriminalIP), and the Uncover-specific engines (Quake, Hunter, PublicWWW, HunterHow, Onyphe, Driftnet). With multiple keys, you can multiply your effective rate limit and avoid scan interruptions from per-key quotas.
Upload and manage custom attack workflow skills (.md files) that teach the agent exploitation techniques beyond the built-in CVE, brute-force, and phishing workflows.
Each skill card shows the skill name, description, and upload date, with actions to edit description, download, or delete.
- Upload Skill (.md) — select a Markdown file, enter a descriptive name and an optional short description (1-2 sentences used by the Intent Router for classification), then click Upload.
- Edit description — click the pencil icon on any skill to update its description without re-uploading the file.
- Delete — removes the skill and automatically disables it in all project configurations.
Uploaded skills appear as toggles in every project's settings, so each project can independently choose which skills to enable.
For details on writing skill files, classification, and built-in skills, see Agent Skills.
Upload and manage on-demand reference skills for the AI agent chat. Unlike Agent Skills above (which drive attack classification and phase-aware workflows), Chat Skills are tactical reference docs -- tool playbooks, vulnerability guides, framework notes -- that you inject into the agent's context on the fly using /skill <name> in the chat.
Each skill card shows the skill name, description, category badge, and upload date, with actions to edit description, download, or delete.
-
Import from Community -- bulk-imports all reference skills from the community folder (
agentic/skills/) into your library. The folder starts empty; click this to populate it with the 46 shipped skills (vulnerabilities, tooling, protocols, technologies, frameworks, Active Directory, cloud, post-exploitation). Skills with the same name are skipped. - Upload Skill (.md) -- upload a custom reference document with name, description, and category.
- Delete -- removes the skill from your library.
Skills uploaded here become available via /skill <name> in all your chat sessions.
For full details on the /skill command, writing Chat Skill files, and the complete comparison with Agent Skills, see Chat Skills.
Curate a personal catalog of trusted security knowledge URLs (HackTricks, PayloadsAllTheThings, CVE PoC repos, vendor research blogs, ...). Each enabled resource becomes a slug the agent can target through the tradecraft_lookup tool during the exploitation and post-exploitation phases.
Unlike web_search (which queries the open web through Tavily) or the FAISS Knowledge Base (which is pre-ingested at install time), Tradecraft is always live and always curated by you. The agent only sees what you explicitly trusted.
For the full lifecycle, type catalog, cache layer, and agent-session flow, see the dedicated Tradecraft Lookup page.
The screen is laid out as:
| Element | Purpose |
|---|---|
| Section header | Title "Tradecraft Resources" with the Add Resource button on the right. |
| Intro line | "Curated knowledge sites the agent consults during exploitation (HackTricks, PayloadsAllTheThings, CVE PoC repos, ...). On add, the agent fetches the homepage, builds a sitemap, and writes a short summary that becomes the tool's catalog entry. The agent only sees enabled resources." |
| Resource cards | One card per saved URL, with the icon, name, type badge, URL link, sitemap entry count, last-verified timestamp, expandable summary, and the per-card action buttons. |
| Field | Meaning |
|---|---|
| Name | Display label (e.g. "HackTricks"). Used to derive the immutable slug (hacktricks, hacktricks-2, ...) that the agent passes as resource_id. |
| Type badge | Auto-detected one of mkdocs-wiki, gitbook, github-repo, cve-poc-db, sphinx-docs, agentic-crawl. While verifying, a verifying… spinner badge replaces it. |
| URL | The base URL clicked at verify time. Click to open in a new tab. |
| Entries | Number of pages indexed in the sitemap (after deterministic build or agentic-crawl). |
| Verified | Relative timestamp (e.g. 2h ago) of the last successful verify or refresh. |
| Summary | Click the expandable text to read the 250-350 word LLM-generated description that the agent will read at runtime. |
| Error | Red error chip with a tooltip when verify failed (HTTP 404, thin homepage, GitHub rate limit, ...). The card stays usable in degraded form. |
| Enabled toggle | Disabled resources are filtered out before the agent's tool docstring is built, so the agent literally cannot call them. Re-enabling re-exposes them on the next project load. |
| Action | Effect |
|---|---|
| Refresh (circular arrow) | Re-fetches the homepage, re-detects the type, rebuilds the sitemap, and rewrites the summary. Use this after a wiki redesign or repo rename. Crawl-type resources will pay the full LLM-driven loop again. |
| Edit (pencil) | Opens the form to change name, URL, GitHub token override, cache TTL, or enabled flag. Saving a URL change re-triggers verify. |
| Delete (trash) | Drops the database row, the cached pages on disk, and the SQLite index entries for that resource. Confirmation prompt required. |
Verify is asynchronous. Adding a resource creates the row immediately so the card appears with a
verifying…chip; the actual fetch + sitemap build + summary runs in the background. The settings page polls every 5 seconds while any resource is unverified, so the badge auto-updates when verify completes.
The modal has two halves: the Quick Add catalog at the top and the resource form below.
A scrollable list of hand-picked, well-maintained, openly-licensed reference sites, spread across all six resource types so users can populate the catalog with one click. Selecting a row fills Name and URL from the preset; you can still tweak them before saving. The chip on the right is just a hint — the backend always re-detects the type at verify time.
The presets are grouped by type below. Use this list to understand which resource type each well-known site falls into:
| Name | URL | Coverage |
|---|---|---|
| HackTricks | book.hacktricks.wiki |
Comprehensive offensive security wiki — web, AD, cloud, privesc, mobile |
| The Hacker Recipes | www.thehacker.recipes |
Pentest methodology by ShutdownRepo — AD, web, infra, exploit-dev |
| CTF Field Guide | trailofbits.github.io/ctf/ |
Trail of Bits CTF guide — vulns, RE, forensics, web, exploits |
| CTF101 | ctf101.org |
Beginner CTF reference — crypto, forensics, RE, web, binex |
| Name | URL | Coverage |
|---|---|---|
| Practical CTF (Jorian Woltjer) | book.jorianwoltjer.com |
CTF + hacking technique notes (web, AD, crypto, binex, mobile) |
| ired.team | www.ired.team |
Red team / offensive security notes — AD, evasion, persistence |
| Name | URL | Coverage |
|---|---|---|
| PayloadsAllTheThings | swisskyrepo/PayloadsAllTheThings |
Payload library + bypass cheatsheets per vuln class |
| InternalAllTheThings | swisskyrepo/InternalAllTheThings |
AD + post-exploitation cheatsheets |
| HardwareAllTheThings | swisskyrepo/HardwareAllTheThings |
Hardware/IoT pentest references — UART, JTAG, BLE, Zigbee |
| h4cker (Omar Santos) | The-Art-of-Hacking/h4cker |
>10k curated hacking references, per-topic folders |
| PEASS-ng | peass-ng/PEASS-ng |
WinPEAS / LinPEAS / MacPEAS privilege-escalation suite |
| SecLists | danielmiessler/SecLists |
Wordlists for usernames, passwords, URLs, fuzzing |
| awesome-pentest | enaqx/awesome-pentest |
Curated meta-list of pentest tools, books, conferences |
| awesome-bug-bounty | djadmin/awesome-bug-bounty |
Bug-bounty programs, writeups, tools |
| awesome-cloud-security | 4ARMED/awesome-cloud-security |
AWS / GCP / Azure cloud security tooling and writeups |
| awesome-web-hacking | infoslack/awesome-web-hacking |
Web pentest tools, books, papers, vulnerable apps |
| awesome-android-security | saeidshirazi/awesome-android-security |
Android security learning path |
| xairy/linux-kernel-exploitation | xairy/linux-kernel-exploitation |
Curated Linux kernel exploitation resources |
| Privilege-Escalation | Ignitetechnologies/Privilege-Escalation |
Linux + Windows privesc cheatsheets |
| API-Security-Checklist | shieldfy/API-Security-Checklist |
Best-practices checklist for testing REST APIs |
| ctf-tools | zardus/ctf-tools |
CTF tool installer collection |
| OWASP CheatSheets | OWASP/CheatSheetSeries |
Concise OWASP cheatsheets per topic |
| OWASP WSTG | OWASP/wstg |
OWASP Web Security Testing Guide source markdown |
| OWASP MASTG | OWASP/owasp-mastg |
OWASP Mobile Application Security Testing Guide |
| cheat.sh | chubin/cheat.sh |
Unified CLI cheatsheets, security tools |
| awesome-iot-hacks | nebgnahz/awesome-iot-hacks |
IoT security resources |
| awesome-malware-analysis | rshipp/awesome-malware-analysis |
RE, sandboxing, YARA, packers |
| awesome-ctf | apsdehal/awesome-ctf |
CTF tools and resources catalog |
| awesome-osint | jivoi/awesome-osint |
OSINT investigation tools and frameworks |
| awesome-shodan-queries | jakejarvis/awesome-shodan-queries |
Curated Shodan dorks |
| awesome-windows-domain-hardening | PaulSec/awesome-windows-domain-hardening |
AD hardening + offensive playbooks |
| awesome-redteam | yeyintminthuhtut/Awesome-Red-Teaming |
Red team operator resources |
| awesome-fuzzing | cpuu/awesome-fuzzing |
Fuzzing harnesses, frameworks, papers |
| Name | URL | Coverage |
|---|---|---|
| trickest/cve | trickest/cve |
Auto-aggregated CVE → public PoC index. Requires cve_id="CVE-YYYY-NNNNN"
|
| 0xMarcio/cve | 0xMarcio/cve |
Alternative CVE → PoC index with extended metadata |
| nomi-sec/PoC-in-GitHub | nomi-sec/PoC-in-GitHub |
Daily-updated CVE PoC index scraped from GitHub repos |
| Name | URL | Coverage |
|---|---|---|
| Scapy docs | scapy.readthedocs.io |
Python packet crafting library — protocols, fuzzing, scapy.layers |
| Volatility 3 | volatility3.readthedocs.io |
Memory forensics framework — plugins, OS profiles, Vol3 API |
| Mitmproxy docs | docs.mitmproxy.org |
TLS-MITM proxy — addons, scripting, intercept replay |
| Sliver C2 | sliver.sh/docs |
Open-source C2 framework documentation (BishopFox) |
| Name | URL | Coverage |
|---|---|---|
| 0xpatrik (Patrik Hudak) | 0xpatrik.com |
Subdomain takeovers, OSINT, recon automation |
| Synacktiv publications | www.synacktiv.com/en/publications.html |
French pentest firm writeups — AD, cloud, mobile, exploit-dev |
| Doyensec research | blog.doyensec.com |
Boutique appsec research — Electron, Java, GraphQL, Solidity |
| SpecterOps | posts.specterops.io |
Red team / AD deep dives (BloodHound team) |
| Project Zero | googleprojectzero.blogspot.com |
Google P0 vuln research — kernel, browser, mobile 0-days |
| Spaceraccoon (Eugene Lim) | spaceraccoon.dev |
Web/cloud/IoT writeups, supply-chain, bug-bounty postmortems |
| Adsecurity (Sean Metcalf) | adsecurity.org |
AD attack/defense — Kerberos, ADCS, replication abuse |
| Trail of Bits blog | blog.trailofbits.com |
Cryptography, smart contracts, fuzzing, OS-level research |
| Tarlogic blog | www.tarlogic.com/blog |
Spanish pentest firm — Kerberos, AD, RT TTPs, threat intel |
| Assetnote research | blog.assetnote.io |
Attack-surface management research and 0-days |
| Field | Required | Default | Notes |
|---|---|---|---|
| Name | yes | — | Free-text label. The server derives slug from this with kebab-case + collision suffix. The slug is stable across renames so in-flight conversations and cache rows do not break. |
| URL | yes | — | The homepage / base URL. Must be http:// or https://. SSRF guard: private, loopback, link-local, multicast, reserved, .local, .internal, and localhost are rejected at verify time. NXDOMAIN returns a separate explicit error. |
| GitHub Token Override | no | (user-level GitHub token) | Per-resource GitHub PAT. Use it when one repo needs different permissions (e.g. an org-private cheatsheet) without changing the user-level token. Hidden by default; click the eye icon to toggle visibility. The dummy hidden username/password pair on the form blocks Chrome / 1Password from polluting this field with a saved login. |
| Cache TTL seconds | no |
0 (use type default) |
Per-URL cache lifetime. Set 0 to inherit the per-type default (7d for wikis/repos, 14d for sphinx-docs, 30d for cve-poc-db, 1d for agentic-crawl). Use a short value for fast-moving blog feeds, a long value for static cheatsheets. |
| Enabled | toggle | true |
When unchecked, the resource is hidden from the agent's tool catalog on the next project load. Verify and refresh still work, so you can keep degraded resources around without exposing them. |
Why no "type" field? Resource type is always auto-detected at verify time from the homepage HTML. A manual override would let users misclassify a site (e.g. mark a generic blog as
mkdocs-wiki), which then breaks the sitemap builder. The chip on each Quick Add row is a hint, not a setting.
The six resource types are not interchangeable — each one drives a different sitemap-extraction strategy, a different cache TTL, and a different at-query-time fetch path. Pick the right type or, equivalently, pick a URL whose homepage will trigger the right detector.
| Type | Detection signal | Sitemap source | Build time | Default TTL |
|---|---|---|---|---|
mkdocs-wiki |
<meta name="generator" content="mkdocs/material"> or <!-- Book generated using mdBook --> or class="mdbook"
|
/sitemap.xml (urlset or sitemapindex) → fall back to mkdocs.yml → fall back to rendered nav harvest |
~5s | 7 days |
gitbook |
host contains gitbook.io, application-name=gitbook, name="generator" content="gitbook", static-2v.gitbook.com, or data-rsc-router
|
/sitemap.xml → fall back to a single Tier-2 Playwright nav harvest |
~10s | 7 days |
github-repo |
host is github.com or raw.githubusercontent.com (and not a CVE repo) |
GitHub Trees API (GET /repos/{owner}/{repo}/git/trees/{branch}?recursive=1), filtered to .md, .txt, .rst files |
~3s | 7 days |
cve-poc-db |
GitHub repo whose name contains cve OR a homepage with >20 CVE-IDs |
Just {owner, repo, branch} — per-CVE lookups are deterministic by ID, so no tree enumeration |
~1s | 30 days |
sphinx-docs |
host ends in .readthedocs.io, or _static/searchindex.js, or name="generator" content="docusaurus", or docusaurus.config
|
searchindex.json (Sphinx) / search-index.json (Docusaurus) parsed for docnames + titles
|
~3s | 14 days |
agentic-crawl |
none of the above matched | Bounded LLM-driven Playwright loop (max 30 pages, 20 LLM calls, 180s wall-clock, depth 3) | 90-180s | 1 day |
-
mkdocs-wiki— for MkDocs / Material / mdBook wikis. They publish a completesitemap.xml, so the build is fast and the sitemap is exhaustive. HackTricks, mdBook-rendered books, and Material-themed docs land here. Multi-language wikis are auto-filtered to English when more than 40% of paths cluster under language codes. -
gitbook— for GitBook-rendered books, including custom domains. They also publishsitemap.xml, but some custom GitBook deployments do not, so a Tier-2 Playwright nav harvest is the documented fallback. ired.team and Practical CTF are typical examples. -
github-repo— for markdown-based knowledge bases hosted on GitHub (PayloadsAllTheThings, awesome-* lists, OWASP guides). The sitemap is the recursive Git tree, filtered to text formats (.md,.txt,.rst). Big repos may be markedtruncatedby GitHub — the resource is then saved with an explicit_errorfield. Anonymous GitHub API is rate-limited to 60 req/h, so adding a GitHub token in API Keys is recommended; the per-resource token override is the per-repo escape hatch. -
cve-poc-db— for CVE-indexed PoC repos (trickest/cve,0xMarcio/cve,nomi-sec/PoC-in-GitHub). The repo has hundreds of thousands of files, so enumerating the tree is wasteful. Instead, the lookup path is built deterministically from the CVE-ID:contents/{year}/CVE-YYYY-NNNNN.md. The agent must passcve_id="CVE-YYYY-NNNNN"; the tool docstring tells the agent so explicitly. Falls back to listing/contents/{year}and substring-matching on miss. -
sphinx-docs— for Sphinx, Read the Docs, and Docusaurus sites. They ship a structuredsearchindex.jsonwithdocnames[]andtitles[], which is the perfect shape for a sitemap. Tool documentation (Scapy, Volatility 3, Mitmproxy, Sliver C2) is the canonical use case. Cache TTL is longer (14 days) because tool docs change rarely. -
agentic-crawl— the fallback for anything that did not match the five deterministic detectors: personal blogs, custom CMS, vendor research portals without a published sitemap. A bounded LLM loop drives Playwright over the site, asking Claude on each page which links to follow next. This is the only type that pays a real LLM cost at verify time (typically $0.30 to $0.60 with Sonnet) and the only type that blocks for ~90-180 seconds on add. Once verified, the sitemap is just JSON in the database and query-time stays sub-second.
A site can change type on Refresh. Detection runs every time. If a wiki migrates from MkDocs to Docusaurus, hitting Refresh on the card swaps the badge from
mkdocs-wikitosphinx-docsand rebuilds the sitemap with the right strategy.
When you start (or reload) a project, the agent reads your enabled resources and dynamically composes the description for the tradecraft_lookup tool. Each enabled resource appears in the docstring as one block:
hacktricks (mkdocs-wiki) https://book.hacktricks.wiki
Comprehensive offensive security wiki. Covers web (XSS, SSRF,
SSTI, deserialization), Active Directory (Kerberoasting, DCSync,
golden ticket, ADCS), Linux/Windows privesc, cloud, container
escapes... [your saved 250-350 word summary]
The whole "which resource fits this query" decision is made by the model reading these summaries. There is no vector router and no learned retrieval. The summary you saved at add-time is the routing intelligence — write good summaries (or trust the LLM-generated default) and the agent will pick the right resource on its own.
When zero resources are enabled, the entry is removed from the tool registry entirely, so the agent does not see a tool that promises a capability it cannot deliver.
The agentic-crawl loop and the cache layer expose four knobs in the project's Project Settings under the agent block (TRADECRAFT_*):
| Setting | Default | What it caps |
|---|---|---|
TRADECRAFT_CRAWL_MAX_PAGES |
30 | Pages visited by the agentic-crawl loop |
TRADECRAFT_CRAWL_MAX_LLM_CALLS |
20 | "Which links to follow?" Claude calls per crawl |
TRADECRAFT_CRAWL_TIME_BUDGET_SEC |
180 | Wall-clock budget per crawl |
TRADECRAFT_CRAWL_MAX_DEPTH |
3 | Max link depth from the homepage |
Plus three runtime knobs:
| Setting | Default | Effect |
|---|---|---|
TRADECRAFT_TOOL_ENABLED |
true |
Master kill-switch for the tool |
TRADECRAFT_FETCH_TIMEOUT |
30 | HTTP timeout (seconds) for Tier 1 / Tier 2 fetches |
TRADECRAFT_TIER2_THRESHOLD_BYTES |
800 | Tier-1 response size below which the tool escalates to Playwright |
TRADECRAFT_DEFAULT_TTL_SEC |
86400 | Fallback cache TTL when both type-default and per-resource override are zero |
TRADECRAFT_SECTION_PICKER_MODEL |
claude-haiku-4-5-20251001 |
Small model used for the at-query-time "which page best answers this?" decision |