Actionbook Browser Agent
Operate logged-in websites and deliver sourced reports, clean CSVs, or ready-to-send drafts.
The README identifies the remote MCP endpoint, Chrome extension, reuse of logged-in sessions, and external actions such as opening tabs, reading pages, clicking, and filling forms, so the broad flow and effect surface are visible. Workflows also use either empty top-level permissions or explicit limited GitHub permissions. Deductions apply because the supplied material does not document extension permissions, what page or credential data the hosted service receives or retains, encryption and deletion practices, per-action confirmation, domain restrictions, sensitive-field protections, or revocation and rollback. Dependency controls include a pinned pnpm release and one override, but most packages use ranges and Actions are pinned only to major tags, with no audit or vulnerability-response policy. Repository branding, community links, and the Cue Labs, Inc. notice provide partial attribution, but publisher identity is unverified and maintenance accountability remains incomplete.
The README, package manifest, and workflows show recognizable build, test, and release structures, including Windows unit and end-to-end configurations; the package manager and ordinary dependency sources are also specified. Deductions apply because the recommended hosted connector and extension implementation are not included in the evidence, the CLI is declared deprecated while release automation still contains CLI installation messaging, and the extension release branch is excluded by the notification job condition. Failure handling is limited to skipping integration tests when an API key is absent, failing a webhook request, and confirming connection during setup; there is no documented user-facing error taxonomy, recovery guidance, or fallback behavior.
The material clearly targets users who need agents to work across authenticated websites, produce reports, CSVs or drafts, and repeat tasks as Recipes. It names several MCP-compatible clients, giving adequate ordinary environment coverage. Deductions apply because capabilities are only summarized through subcommands such as search, manual, goto, snapshot, click, and type, while the complete command set, supported sites, read/write boundaries, and concurrency limits are deferred to external descriptions. Trigger guidance consists of a natural-language example and an optional explicit prompt, without statically reviewable routing rules, ambiguity handling, or safeguards against triggering on unsuitable pages.
The README has clear sections for positioning, installation, quick start, tools, documentation, community, contribution, and licensing. Its three-step hosted-connector setup is adequate for ordinary installation. The complete Apache-2.0 text agrees with package metadata, justifying full license marks. Changesets, version synchronization, release scripts, and tag conventions show a release path, but no changelog content is supplied and the root package remains version 0.0.0. Deductions also reflect the coexistence of the current product, a Legacy section, a deprecated CLI, and continuing CLI release logic, which weakens naming and lifecycle clarity. Examples are sparse, no FAQ is present, limitations and support coverage are largely external, and no specific maintainer or responsibility policy is identified beyond general community channels.
The README describes directly usable deliverables including sourced reports, clean CSVs, ready-to-send drafts, and reusable Recipes. Access to authenticated data, API-first operation with UI fallback, and concurrent tabs plausibly offer incremental value over generic browser agents. Deductions apply because claims of 10x speed, accuracy, resilience, and processing 192 sites in two minutes are marketing statements or a video reference without methods, baselines, or result data in the supplied files. Installation is described as free and CLI-free, but hosted-service pricing, privacy costs, service limits, resource consumption, and maintenance burden are not disclosed, leaving cost-benefit support thin.
Some claims trace to concrete artifacts, including the Apache license, pnpm version, test workflows, hosted endpoint, and CLI deprecation notice. The README, package manifest, and workflows also corroborate the project identity, package structure, and release machinery. Deductions apply because the core agent, hosted connector implementation, extension manifest, privacy controls, comprehensive tests, and linked documentation are absent, preventing closed-loop verification of the principal security and performance claims. Marketing assertions, demonstration results, and architectural implications are not systematically separated from verified facts, so all three criteria receive only thin support.
- The product reuses live Chrome login sessions and can click and fill forms. Before granting access, independently review extension permissions, remote data transfer, retention and deletion practices, and confirmation controls for consequential actions.
- Do not treat the 10x faster, accurate, or 192-sites-in-two-minutes claims as independently verified; the supplied material contains no benchmark method or auditable results.
- The recommended product is a hosted MCP connector plus Chrome extension, while the supplied source mainly exposes repository metadata, legacy CLI-oriented automation, and limited test configuration; the runtime implementation cannot be fully audited here.
- The CLI deprecation notice coexists with release automation that still emits CLI installation text. Confirm the current support path, version compatibility, and recovery procedures before adoption.
What does this agent do, and when should you use it?
Actionbook is a browser agent that runs in Chrome’s side panel through a remote MCP connector paired with a Chrome extension. It reuses sessions in the user’s real browser to read data behind logins or paywalls and operate websites already in use. Its execution model favors direct API requests when possible and falls back to UI automation when necessary, while a stateless architecture enables concurrent work across tabs. Outputs can include sourced reports, clean CSV files, and ready-to-send drafts rather than conversational answers alone. Successful workflows can be saved as Recipes and rerun with different inputs. The documented deployment boundary depends on Actionbook’s hosted MCP endpoint and Chrome extension; no self-hosted path is provided.
The user connects an MCP-compatible AI application to https://edge.actionbook.app/mcp and links a real Chrome session through the Actionbook extension. The agent then calls the single actionbook MCP tool, using search and manual to retrieve action manuals and browser-control subcommands such as goto, snapshot, click, and type to open pages, inspect snapshots, activate controls, and fill forms. It can reuse authenticated browser sessions and execute work across multiple pages or tabs concurrently. Website operations use direct API calls where possible and UI automation as a fallback. The resulting work can be delivered as a sourced report, clean CSV, or send-ready draft, and repeatable tasks can be stored as Recipes.
- An operations specialist who needs information from several authenticated SaaS products can ask it to open services such as Stripe, Linear, and Vercel and collect the required data.
- A researcher working across login-gated sources can reuse an existing Chrome session and request a sourced report rather than manually copying each page.
- A research team checking many company websites can use concurrent tabs instead of processing every site sequentially.
- A team with a recurring browser workflow can save a successful task as a Recipe and rerun it with new inputs.
- A business user who needs a practical handoff can request a clean CSV or ready-to-send drafts instead of receiving only a prose answer.
What are this agent's strengths and limitations?
- It reuses authenticated sessions in the user’s real Chrome browser, allowing work on pages behind logins without documenting a separate authentication integration for every site.
- Direct API requests are combined with UI automation fallback, reducing reliance on page-by-page visual interaction when an API path is available.
- Action manuals provide concrete operating instructions; the README claims this approach can be 10 times faster than repeated snapshot parsing.
- Its stateless architecture supports concurrent operation across many tabs, which fits bulk website research.
- It produces usable deliverables such as sourced reports, CSVs, and drafts, and successful workflows can be saved as Recipes.
- Core execution depends on Actionbook’s hosted
https://edge.actionbook.app/mcpendpoint, and no self-hosted deployment is documented. - The full workflow requires Google Chrome and the Actionbook extension, so it is not presented as a browserless server-only runtime.
- Because the extension and remote connector operate authenticated pages, adopters must evaluate access to sensitive accounts and page data.
- The standalone
@actionbookdev/cliis deprecated, creating migration work for users of the legacy CLI model. - The source does not document pricing details, service guarantees, a supported-site matrix, or precise behavior when both API access and UI automation fail.
How do you install or deploy this agent?
No CLI installation is required. First, install the Actionbook extension from the Chrome Web Store. In Claude, ChatGPT, Perplexity, Hermes, OpenClaw, or another MCP-compatible application, open the connector settings, add a custom connector, and enter https://edge.actionbook.app/mcp. Then sign in at https://actionbook.app/dashboard, follow its client-connection flow, and confirm that the extension is linked. The documented runtime requires Google Chrome, network access, the Actionbook extension, and an AI application capable of connecting to a remote MCP endpoint. The standalone @actionbookdev/cli package is deprecated and should not be used for a new setup.
How do you use this agent?
After linking the extension and MCP client, give the agent a direct task such as: Open stripe.com, linear.app, and vercel.com, then grab the pricing from each. Actionbook retrieves action manuals and drives the real Chrome session by opening tabs, reading snapshots, clicking, and filling forms. To invoke it explicitly, add: Use Actionbook to understand and operate the web page. Once a workflow behaves as intended, save it as a Recipe and rerun it later with different inputs.
How does this agent compare with similar options?
The README contrasts Actionbook with browser agents that repeatedly capture and parse a page before deciding on every next step. Actionbook instead uses action manuals, attempts direct API calls first, and falls back to UI automation, while its stateless architecture allows concurrent tabs. It claims a 10× speed advantage, but the supplied material contains no independent benchmark and names no specific competing product.
FAQ
Must I give Actionbook my website passwords?
Can I self-host the MCP service?
https://edge.actionbook.app/mcp; it does not provide a self-hosting procedure.Can I keep using @actionbookdev/cli?
What happens when a site cannot be accessed through a direct API?
snapshot, click, and type. The source does not claim that every website or interface is supported.