Forage

Put web research in every agent's hands and keep what it found in a per-thread Research tab you can read, quote and cite. web_search, web_fetch, web_map, web_extract, web_crawl and web_diff run as bb's own tools for every provider, with routing profiles, domain and date filters, bounded limits and every partial failure in words. Replies carry citation cards that quote the exact kept lines a claim rests on, and Check in sources matches any message against the pages stored for that thread. A zero-turn CLI runs the same operations; private networks are blocked by default and keyless DuckDuckGo and Exa search need no account.

Overview

View source

Forage gives every BB agent six web-research tools, durable scoped evidence and a CLI that uses no model turns. Requires BB 0.47 or newer.

Research tools

web_search: up to eight queries, domain/date filters, provider profiles and fair host-diverse content reads.
web_fetch: up to ten URLs, readable or exact textual raw mode, independent direct/GitHub-smart resolution, passage search and outline/link/line/focus views.
web_map: bounded sitemap, robots, llms.txt and HTML discovery with provenance.
web_extract: deterministic metadata, headings, links, tables and JSON-LD. No hidden AI extraction.
web_crawl: persistent start/status/list/cancel/resume jobs with page, depth, byte, time and concurrency limits.
web_diff: immutable snapshot comparison or a fresh URL against stored evidence.

Content is captured before formatting. Continue with snapshotId, contentHash and UTF-16 offset for immutable paging. URL-only paging may download again after cache expiry. Every batch failure and partial limit is visible; output is capped at 32,000 characters.

Research in each thread

A Research tab in each thread shows what the agent and you looked up: searches with the results seen, pages read, crawls against their limits, and problems in words. Read kept text, compare versions, quote lines into your message, check a reply against kept pages, and search or crawl yourself with no model turn. Replies can carry citation cards. The thread menu sets web research per thread: Docs, Fresh, Private or Off. The Research page lists threads with research, plus health and storage.

The activity log keeps queries and results for 30 days; clear it per thread or turn it off. Source IDs stay in their thread unless shared with the project.

Default retention is 30 days for thread content, 180 days for project content and 128 MiB retained content. Active-job snapshots are protected. Storage refusal never produces a fictional source ID. Missed deletions are recovered by scope checks; archived threads are kept.

Privacy and providers

Keyless DuckDuckGo and Exa work without credentials. Optional Brave, Exa account, SearXNG and Tavily search, explicit Tavily/Firecrawl/Jina readers, and an explicit Firecrawl crawl backend respect configured ceilings. Keys do not automatically enable paid fallback.

Profiles: auto, fast, fresh, docs, private and deep. Private allows configured SearXNG search, direct readers and local crawl only; it makes no guarantee about SearXNG's upstream engines. Per-call arguments cannot widen setting-level privacy.

Private networks are blocked by default and each direct redirect is revalidated and DNS-pinned. Local private-network permission is not permission to forward URLs to a cloud reader. Remote URLs are screened before forwarding; vendor-side redirects cannot be DNS-pinned locally.

Restricted same-thread capture is OFF by default, cannot be project-promoted, and needs explicit thread selection for export. Disabling it revokes reads. Credentials/headers/cookies are never persisted. Sensitive URL checks are conservative heuristics, not a general secret detector. Scope isolation does not protect SQLite from a local user or shell-capable agent.

CLI and diagnostics

Sources commands: list, show, export, delete and prune. Destructive commands require exact confirmation tokens; export returns a local-auth download route, not a server file path.

status and diagnostics show redacted in-memory records (maximum 200, 24 hours), never query/URL/content/credential values. clear clears process caches. doctor is offline unless --live explicitly enables bounded keyless probes. bench runs a bundled local corpus without network access.

See the README for every parameter and limit.

More from Kalindu De Costa

BurnSee what every agent thread costs in dollars while it runs, per prompt and across all your threads. A chip beside the context circle shows the thread's total and what the last prompt added; click it for where the money went, a chart of every model call and the total with child threads. The Burn tab in the side panel lists the cost of each prompt, and a Burn page in bb's navigation charts spend across threads by day, project, model and harness. Budgets notify you at 80% and 100% of a thread's budget, and stop a run only if you turn that on. Works with pi, OpenCode, Codex, Claude Code and any harness that reports usage to bb; where a harness records no cost, Burn estimates it from list prices and marks it with ~.165 LoomHand a goal to agents and keep the decisions. A planner turns it into a spec with tickets, acceptance criteria, checks and a merge plan, and a separate agent reviews the spec before you approve it. Workers then build each ticket in its own git worktree, Loom runs your checks, and reviewers on another model approve the work or send findings back. Approved work merges into an integration branch you land as one pull request, or opens as one pull request per ticket that Loom follows through CI. A full-screen Loom page shows what needs you, the board, every attempt and review, the changes before they land, insights on how the project went and a timeline you can rewind.New Session LogCheck what an agent actually ran, changed and decided, in one live log beside the chat. Messages, thinking, commands with their output and exit codes, tool calls, file diffs and subagent steps appear in timeline order, with a timestamp on every row and a duration on every command. The log follows new output as it arrives, opens failed commands by default, and filters to messages, thinking, tools or errors. Search covers the whole session, and single keys jump between rows, replies and errors. Open it with Mod+Shift+L or from the thread header; it reads bb's timeline and keeps no copy of the thread.New ShuttlePuts every pull request that needs you, from any server, in one place. The Inbox lists the decisions waiting on you, your PRs keep the agent thread that fixes their reviews and failed checks, and the review centre groups PRs that ask you or your teams by where your review stands. Merge, comment and post AI review drafts from Shuttle, on GitHub out of the box or on any other server you describe with a command adapter. Insights show where review time goes: who reviews your PRs, how fast, and what your agents fixed.New