Changelog

v3.71

Choose which CSV rows a batch runs

Batch creation now previews the CSV you upload and lets you pick the rows to run. The results grid lists every row, with compact cards and load-more for large batches.

v3.68

Evaluate agents where you build them

When you create an agent, Grep writes a first set of evaluation questions, and you edit them to say what the agent has to get right. Your team can then score the agent's work as it runs and leave feedback on its accuracy, next to the output rather than in a separate spreadsheet. That feedback flows back into the agent, and Grep proposes prompt updates you review and apply one by one, the way you would a pull request.

Play release video

v3.67

Connect the tools your team already uses

Agents can now work inside your own accounts, not only Grep's built-in data integrations. Connect SaaS accounts such as Gmail, Outlook, Notion and Salesforce, add your own custom APIs and MCP connections, and for internal tools that have no API, sign in through the browser and let the agent use the application. Credentials can be kept in an encrypted vault, or you can sign in by hand each time the session expires.

Play release video

v3.66

Research that ends in a decision

Choose Decision as an agent's output type, write the questions, the answer options and a rubric for each, and once the research is done TypeSafe AI's Jev model decides against those criteria. For choice questions it returns a calibrated probability per option, so you can set a threshold; above it the agent acts, below it a person reviews the case. The decisions sit above the supporting report and download as JSON.

Play release video

v3.64

Turn a procedure into a reusable skill

Describe a procedure, attach the documents your team relies on, and Grep turns it into a skill in your Skills Library. Use that skill in chat or add it to any agent that needs the procedure, and when the procedure changes, update the skill once and every agent that uses it follows the new version. In the launch demo, one merchant-classification skill serves both an enterprise and a small-business onboarding agent.

Play release video

v3.62.1

Catalog agents now deliver finished files

Twenty-four catalog agents now hand you a finished file, a spreadsheet, a Word document, a slide deck, a dashboard or a PDF, instead of a Markdown report. Home now gathers your agents, running batches and recent reports in one place, and a chat reply can hand over a workspace file as a direct download. Developers get a shorter MCP address and tools named agent_ and run_; the old names keep working.

v3.43

Every account gets Loops, monitors and batches

Loops, monitors and research batches now work on every account, a Loop can run every hour, and a batch can hold up to 1,000 rows. When you read someone else's public run, Get this weekly copies that agent into your account and starts a Loop of your own. API keys and their scopes now live in Settings, and new personal accounts start with 500 Wallet credits.

v3.24.1

Direct uploads from the API and MCP

A file of up to 25 MB can now go straight to Grep from the API or MCP, without its bytes passing through an agent's context. A run can drive up to four named browsers at once, each with its own card in the job view, and people you share a private run with can follow its activity. Dark mode is on for every account, and any account can request a Team with a shared wallet and invitations.

v3.14

Watch your agent's browser at work

A run's Work journal now lists every browser session it opened. You can watch the live view while the run is going, and afterwards the recording plays back inline, with the session's tabs, proxy details and last page. Chat artifact cards show the run's plan as it works and list every file it delivered.

v3.10

Filter agents and export job history

The Agents page now filters by status, owner and tag, and anyone who edits an agent can add shared tags. Job history sorts and filters on the server and exports to CSV, up to 10,000 rows at a time. An agent's configured effort is prefilled wherever you start a run from.

v3.9

Homepage starters become public agents

The starter agents on the homepage are now public agents, so anyone can open one, fill in its form and run it, signing in only when they submit. Every browser session an agent opens now runs with proxies and advanced stealth, so sites that turn away automated visitors are easier to reach.

v3.6.7

A public address for every published report

Published research now has a canonical public address that opens quickly for readers who are not signed in, and Word, PowerPoint, Excel and PDF publications open in the browser. A new sidebar puts the team switcher on every page and adds a Chats list and a Cmd+K command palette. Experts are now called Agents, and developers can create, build and publish an agent entirely through the API and MCP.

v3.4.2

Shared reports get a shorter way in

Someone who opens a report you shared now sees a small prompt with a single Continue button, and the report blurred behind it. Closing the prompt leaves a slim banner for the rest of the session instead of bringing it back on the next page load. Once they sign up they land straight back on the report; the profile questions can wait.

v3.1.1

Set a spending cap on each run

A research request sent through the API can now carry a maximum cost for that run, from $0.50 to $1,000, and an agent can declare a cap of its own; when both are set, the lower one applies. PowerPoint outputs fill the exact template you supply, and past invoices are listed in Billing.

v3.0.3

Document verification reads every file type

Document verification now accepts every file type you upload, images included. A scanned page or a photo in any raster format is read as an image when the file holds no extractable text, so it gets the same checks as a text document.

v3.0

Introducing Agent Builder

Build, improve, and evaluate your own Agents in one place: a redesigned editor with output templates, type-aware HTML and file editing, skill and tool previews, reorderable inputs, and international address autocomplete, plus per-session chats, an activity inspector, and context files on every run path.

v2.9

Sign in anywhere and clearer completion emails

Enchanted links now work more reliably across devices, while research and build completion emails lead with readable titles, content, artifact labels, and direct actions.

v2.8

Richer monitor and schedule notifications

Monitor and scheduled-report emails now include useful report content, newspaper-style daily briefings, focused previews, and threaded lifecycle alerts.

v2.7

Stronger source confidence and broader context files

Reports gain source-confidence rails, verified links, and licensed paywalled-news retrieval, while the Expert builder accepts presentations, spreadsheets, and images up to 25 MB.

v2.6

Customer context and research safeguards

Research owners get a customer Info tab, clearer work journals and evidence separation, and stronger anti-hallucination safeguards across research, builds, and widgets.

v2.5

Redesigned, confidence-aware reports

Reports receive clearer typography, numbered sections, mobile headers, verdict stamps, source checking, section-level confidence, screenshot-backed URL verification, and usage-based hours saved.

v2.4

Streaming reports, unified monitors, and messaging

Research reports stream with faster structure and richer citations, monitors and schedules share a unified flow, and Telegram and WhatsApp support expand how agents deliver results.

v2.3

Custom expert skills and expanded sharing

Per-expert SKILL.md instructions now reach research and chat runs, the Expert builder gains icon and model controls, and expert sharing, copying, editing, and document verification expand.

v2.2

Paid research and safer web workflows

API research gains x402 and MPP payments with automatic failure refunds, safer outbound URL fetching, podcast intelligence workflows, Markdown reports, and reliable effort selection.

v2.1

MCP, A2A discovery, and research continuations

A public MCP server and A2A discovery make research agents easier to use, while continuations, reference jobs, public research pages, and richer sidebar job details broaden the workflow.

v2.0

Grep 2.0

A new public API brings structured outputs, timelines, files, and a rebuilt playground alongside skills-first experts, scoped API keys, spend controls, and multi-channel chat foundations.