Docs: pair Delta Browser with Jiro Search (search-then-verify)

README.md

@@ -118,10 +118,27 @@ Delta Browser is a single binary that sits between those two:
 
 ---
 
+## Pair it with Jiro Search
+
+Search is only half of a research agent. The other half is opening the source,
+reading it, and checking the claim before acting on it.
+
+**Jiro Search** (open-source, self-hosted) returns cited claims instead of
+links. **Delta Browser** opens those sources, snapshots the page, and extracts
+the exact text — so an agent can verify a claim instead of trusting a snippet.
+
+See [docs/JIRO-INTEGRATION.md](docs/JIRO-INTEGRATION.md) for the full
+search-then-verify workflow, MCP config for both servers, and the safety
+limits (separate browser profile for untrusted pages, sandbox only on Linux
+x86_64).
+
+---
+
 ## Table of contents
 
 - [What is Delta Browser?](#what-is-delta-browser)
 - [Why an agent needs its own browser](#why-an-agent-needs-its-own-browser)
+- [Pair it with Jiro Search](#pair-it-with-jiro-search)
 - [Install](#install)
   - [Linux and macOS](#linux-and-macos)
   - [Windows](#windows)
@@ -692,6 +709,7 @@ document or URL changed. Take a second `snapshot { delta: true }`.
 | [Driving from an agent](docs/AGENTS.md) | The navigate-snapshot-act loop, batching, safety |
 | [Agent skill](docs/AGENT-SKILL.md) | The copy embedded in the binary, for agents to read |
 | [Comparison](docs/COMPARISON.md) | Delta Browser vs Playwright MCP vs Puppeteer MCP |
+| [Jiro integration](docs/JIRO-INTEGRATION.md) | Search with Jiro, verify with Delta Browser |
 | [FAQ](docs/FAQ.md) | Long-form answers to the common questions |
 | [Migration](docs/MIGRATION.md) | Coming from AeroBrowser 0.5 |
 | [Architecture](docs/ARCHITECTURE.md) | System design |

docs/JIRO-INTEGRATION.md

@@ -0,0 +1,79 @@
+# Search with Jiro, verify with Delta Browser
+
+Jiro Search and Delta Browser solve two different halves of the same job.
+
+- **Jiro** searches and returns **cited claims** instead of a pile of links.
+  It is open-source, self-hosted, and has 22 MCP tools, 9 search engines and
+  12 platforms. See `github.com/DevAnimecx/jiro` — `pip install jirosearch`,
+  then `jiro serve`.
+- **Delta Browser** opens a real browser, snapshots the page, and extracts
+  the exact text, so an agent can check a claim against its source before
+  acting on it. It has 62 MCP tools, batched calls, and delta snapshots.
+
+Neither tool replaces the other. Jiro finds and cites. Delta opens and proves.
+
+## The workflow
+
+1. **Search in Jiro.** Ask for claims, not links, and keep the source URL
+   attached to every claim.
+2. **Open the source in Delta.** Use Delta's `navigate` tool with the
+   source URL from step 1, not a URL the model guessed.
+3. **Snapshot, don't screenshot-read.** Use Delta's `snapshot` with
+   `interactive_only: true` first, then `get_text` for the exact passage.
+   Screenshots are for a human to eyeball; text is what an agent should
+   quote back.
+4. **Quote or reject.** If the page text supports the claim, quote it with
+   its source. If it does not, say so and do not act on the claim.
+   A claim that fails step 3 is not a fact, even if Jiro returned it.
+5. **Batch the boring parts.** Where Delta supports it, use its `batch`
+   tool to run navigate + snapshot in one MCP message instead of one
+   round-trip per tool call.
+
+## Example flow (illustrative, not a live run)
+
+```
+User: "What is Groq's free-tier request limit, with a source?"
+
+Agent (Jiro):  returns a claim — "Groq: 1,000 requests/day" —
+               cited to Groq's pricing/docs page.
+Agent (Delta): navigate -> that source URL
+               snapshot { interactive_only: true }
+               get_text -> the passage stating the limit
+Agent (user):  "Groq: 1,000 requests/day — source: <that page>,
+               passage quoted above. Verified in the browser, not
+               just from a search snippet."
+```
+
+If Delta cannot load the page, or the passage is missing, the agent must
+report that failure instead of repeating the Jiro claim as verified.
+
+## MCP setup (both servers, same agent)
+
+Run each server the way its own README documents it — Delta via
+`delta mcp` (see `docs/QUICKSTART.md`), Jiro via its own MCP server
+(see the Jiro README). Do not merge their configs into one server:
+they are separate processes with separate tool namespaces.
+
+In your agent's MCP config, add both servers, then restart the agent.
+Verify with Delta's `delta doctor` and with Jiro's own health check
+before giving the agent a real research task.
+
+## Safety limits — read these before you automate
+
+- **Separate profile for untrusted pages.** Delta inherits the browser
+  profile and login state of the browser it drives. Do not point it at
+  your main logged-in profile to scrape untrusted sites.
+- **Sandbox is Linux-only.** Landlock + seccomp + cgroup v2 isolation
+  applies on Linux x86_64 with the managed engine. On Windows and macOS
+  there is no kernel isolation, and Delta's `server_status` says so.
+  Do not promise sandbox protection on those platforms.
+- **No shared auth.** Jiro and Delta do not share credentials. A Jiro
+  API key, if any, is not a Delta credential, and vice versa.
+- **Do not auto-act on unverified claims.** Filling a form, sending a
+  message, or spending money on the basis of a search snippet alone is
+  exactly the failure this pairing is meant to prevent. Verify first,
+  then act, and log the source you verified against.
+- **Rate limits still apply.** Opening many source pages in a loop can
+  hit a site's limits or look like scraping. Batch sensibly, respect
+  robots/terms for the target site, and stop on blocks instead of
+  retrying aggressively.