Tavily, Firecrawl, and Cloudflare providers
Three new providers join the lineup: Tavily and Firecrawl add web search and contents support, while Cloudflare Browser Rendering enables full-page extraction for JavaScript-heavy sites. The web research workflow is now fully asynchronous, and the provider settings view shows each backend's supported capabilities at a glance.
💥 Breaking changes
Async web research workflow
The web_research tool now always runs asynchronously and uses one execution model across providers.
Start research as before:
{
"input": "Compare the managed cloud SIEM market in 2026"
}pi now returns immediately, tracks the running job, and later posts a completion message with the saved report path.
If you previously passed research-specific local execution controls such as requestTimeoutMs, retryCount, retryDelayMs, pollIntervalMs, timeoutMs, maxConsecutivePollErrors, or resumeId in web_research.options, remove them. The async workflow is now the only supported research behavior.
You can still configure the overall async research deadline in ~/.pi/agent/web-providers.json with settings.researchTimeoutMs or a provider-specific override such as providers.gemini.settings.researchTimeoutMs.
Legacy provider config cleanup
Legacy provider-local enablement fields are no longer accepted in ~/.pi/agent/web-providers.json.
If your configuration still uses providers.<id>.enabled or providers.<id>.tools, move capability routing to the top-level tools mapping instead.
Before:
{
"providers": {
"exa": {
"enabled": true,
"tools": {
"search": true,
"contents": true
}
}
}
}After:
{
"tools": {
"search": "exa",
"contents": "exa"
},
"providers": {
"exa": {
"apiKey": "EXA_API_KEY"
}
}
}The release also removes the legacy custom web_contents response variant and continues a broader simplification pass across configuration parsing, provider wiring, diagnostics, and the in-memory contents cache. Background prefetch, per-URL reuse, TTL expiry, and session reset behavior still work the same way, but the internals now use smaller, clearer abstractions with less compatibility and caching cruft.
By @mavam.
🚀 Features
Cloudflare Browser Rendering provider for web contents
The new cloudflare provider adds web_contents support via Cloudflare Browser Rendering. It renders pages in a real browser and converts them to Markdown, which helps with JavaScript-heavy sites that do not expose useful content to a plain HTTP fetch.
Configure it with a Cloudflare API token and account ID:
{
"tools": {
"contents": "cloudflare"
},
"providers": {
"cloudflare": {
"apiToken": "CLOUDFLARE_API_TOKEN",
"accountId": "CLOUDFLARE_ACCOUNT_ID"
}
}
}You can also pass Browser Rendering options such as gotoOptions, waitForSelector, waitForTimeout, and cacheTTL through providers.cloudflare.options or per-call web_contents.options.
By @arpagon, @mavam, and @codex in #9.
Firecrawl provider for web search and contents
The new firecrawl provider adds web_search and web_contents support via Firecrawl's official JavaScript SDK.
Configure it with a Firecrawl API key:
{
"tools": {
"search": "firecrawl",
"contents": "firecrawl"
},
"providers": {
"firecrawl": {
"apiKey": "FIRECRAWL_API_KEY"
}
}
}You can pass Firecrawl search options such as sources, categories, location, timeout, and scrapeOptions through providers.firecrawl.options.search or per-call web_search.options.
You can pass Firecrawl scrape options such as formats, onlyMainContent, waitFor, headers, location, mobile, proxy, and cache controls through providers.firecrawl.options.scrape or per-call web_contents.options.
Tavily provider for web search and contents
The new tavily provider adds web_search support via Tavily Search and web_contents support via Tavily Extract.
Configure it with a Tavily API key:
{
"tools": {
"search": "tavily",
"contents": "tavily"
},
"providers": {
"tavily": {
"apiKey": "TAVILY_API_KEY"
}
}
}You can pass Tavily Search options such as searchDepth, topic, timeRange, includeRawContent, and exactMatch through providers.tavily.options.search or per-call web_search.options.
You can pass Tavily Extract options such as extractDepth, format, includeImages, query, and chunksPerSource through providers.tavily.options.extract or per-call web_contents.options.
🔧 Changes
Provider capability summary in /web-providers
The /web-providers settings view now shows each provider's supported capabilities directly in the Providers section, so you can see at a glance which backends support web_search, web_contents, web_answer, and web_research without leaving the UI.
For example, the provider list now includes capability columns alongside the provider status:
Providers
Provider S C A R Status
Exa ✔ ✔ ✔ ✔ configured
Codex ✔ builtin
This makes it easier to choose tool mappings without having to cross-reference the README.
🐞 Bug fixes
Claude and Codex no longer preflight local auth
The Claude and Codex providers no longer try to preflight local login state before exposing their managed tools.
Instead of probing CLI auth state up front, the extension now only validates obvious structural setup such as an explicitly configured executable path that does not exist. Real authentication failures are now surfaced by the underlying CLI at runtime.
For Codex, this also removes the old checks for environment variables and ~/.codex/auth.json, which could produce false negatives for setups that use Codex's OS credential store, custom CODEX_HOME paths, or wrapper-based workflows.
Gemini research reliability and reporting
Gemini-backed web_research runs now fail earlier with clearer diagnostics when the provider stops making forward progress, and the resulting completion messages are easier to review in pi.
The overall research deadline is now configurable through settings.researchTimeoutMs or a provider-specific override such as providers.gemini.settings.researchTimeoutMs. The default deadline is now 30 minutes.
When Gemini returns terminal states such as requires_action or incomplete, the extension now reports those outcomes directly instead of continuing to poll until the outer timeout expires.
Async research result messages also render more cleanly in the TUI. The collapsed view now shows a compact metadata summary with the start time, duration, saved file, and any error, while ctrl+o expands the full report. Successful expanded reports now render as Markdown.
