Browserr 0.7 brings Servo pages to the terminal
By AgentRiot Editorial
Browserr adds native terminal rendering and named MCP engines, then fixes a preview bug that could send agents into needless rewrite loops.

Browserr 0.7 puts rendered web pages inside a terminal pane, with mouse input, scrolling and tabs. Its other substantial addition is less visible: an MCP client can now keep several named browser engines running, each with its own proxy and user-agent settings. Both features arrived in 0.7.0 on November 15. The 0.7.1 release (opens in new tab), published later that day, corrects a field-reading bug that could make an agent believe a successful long write had been truncated.
These changes address different parts of an agent's browsing loop: seeing the page, keeping browsing contexts separate, and checking whether an action worked. The changelog (opens in new tab) makes the release split explicit. The terminal and named sessions are the 0.7 update; the preview correction is the patch.
Browserr uses a Rust engine built around Servo 0.8, not Chromium or the Chrome DevTools Protocol. Servo (opens in new tab) is an embeddable rendering engine; Browserr supplies the automation protocol, Python SDK and MCP server around it. The tagged README (opens in new tab) describes 42 MCP tools covering page reads, form interaction, screenshots and handoff helpers. Rendering defaults to CPU, with --gpu available for hardware rendering. Choosing another engine also means accepting another engine's compatibility limits.
Pixels in the terminal, not a text-only page
Browserr Term displays the engine's rendered page rather than turning HTML into a text-browser layout. Sixel and Kitty graphics provide native-resolution output in compatible terminals. A half-block renderer supplies a lower-resolution alternative using colored character cells. Mouse clicks and wheel events go back to the page; keyboard shortcuts manage navigation and tabs.
The important implementation change is where the image work happens. The existing screenshot API returns a full BGRA image. The new terminal path instead asks the engine for a frame. Rust reads Servo's RGBA framebuffer, compares it with the previous image for that tab and stream, and encodes changed rectangles as PNG or terminal-ready Sixel. The frame-pipeline documentation (opens in new tab) describes the new FRM2 packets and their separate per-stream reference images.
Rectangles align with terminal cells. Adjacent changed regions can be merged, while navigation and viewport changes invalidate the old image and require a full frame. Each request ends with an explicit empty packet, including the unknown-tab path, so the client can distinguish completion from waiting for more patches. That behavior is visible in the engine implementation (opens in new tab).
For a page with a small changing counter, the design can avoid transferring a complete raw screenshot on every refresh. It cannot make a scrolling page or full-screen animation stop changing. The mechanism reduces redundant pixel transfer; it does not establish a measured latency or frame-rate advantage.
The two image paths also behave differently. Sixel patches can be written at their terminal coordinates, but color quantization means gradients lose fidelity even when text retains native resolution. In Browserr's Kitty path, PNG patches update a canvas in Python and the composed image is retransmitted when it changes. That path is not patch-only traffic all the way to the terminal.
There are practical prerequisites beyond image support. The terminal guide (opens in new tab) lists Python 3.12; the Kitty compositor needs Pillow, and the half-block fallback uses NumPy. Native Sixel avoids those Python image-processing dependencies. Inside tmux, graphics need passthrough and a cell-size override such as --cell 16:32, adjusted to the actual font. Terminal resizing currently requires restarting the client. “Works in a terminal” therefore needs more qualification than “runs wherever Python runs.”
A named session owns an engine
A session name is more than another tab label. Proxy and user-agent settings are engine-wide properties in this integration. To browse with different settings concurrently, Browserr starts separate engines and routes later tool calls to the requested session.
The MCP server (opens in new tab) keeps a map of session names to drivers. Newly spawned non-default sessions receive separate ports. Calls without a session retain the default browsing context. An already-running engine cannot have its proxy, user agent or headed mode changed by attaching with different arguments; those choices belong to engine startup.
The shipping MCP contract is incomplete, however. Only browser_start and browser_stop advertise session in their input schemas. The dispatcher accepts it for most other tools, but a schema-guided client is not told to supply it. The committed session test (opens in new tab) calls that dispatcher directly; it does not establish that ordinary clients will discover the routing correctly.
There is also a handoff exception in the tagged MCP server (opens in new tab): browser_wait_for_user reads the default driver rather than the selected named session. With a default engine it can watch the wrong browser; without one it errors. A multi-session client must not assume that this helper confirms completion in the browser it paused.
Persistence is a separate choice. By default, spawned sessions use ephemeral profiles; passing a profile selects persistent storage so logins can survive. A session name routes calls, while a profile name selects stored state. Developers who need separate identities should use separate profiles rather than assume different session labels make a reused profile private.
This is useful process and state separation, not an operating-system security sandbox. It also means each additional spawned session carries another engine's resource cost. Browserr provides no basis here for claiming that arbitrary numbers of concurrent agents are cheap.
The 0.7.1 fix makes a preview identify itself
Before the patch, read_structure() clipped reported text-field values to the first 200 UTF-16 code units without saying so. An agent could write a long comment successfully, read the field structure to check its work, see only the beginning and retry a write that had already succeeded. The defect was in the description of the field, not necessarily in the stored text.
Version 0.7.1 keeps that short preview but adds value_clipped when clipping applies. The name can be misleading: it contains the longer string's JavaScript length, not a Boolean. The driver implementation (opens in new tab) still slices the reported value to 200 and records its length separately.
An agent should interpret that metadata as “this is a preview,” not “the page lost the rest.” A matching length alone does not prove two long strings are identical. Workflows that require exact verification still need the full field value or another application-level confirmation before submitting or retrying.
The committed long-value regression fixture (opens in new tab) checks agreement between the write result, the preview and its reported length. The separate question for an agent is whether the remote application accepted the final submission.
A form-filling agent can use the existing helpers to complete that loop: read the structure, identify a field, write through set_field, inspect the result, then submit through submit_form and read the outcome. The form API (opens in new tab) reports the resulting URL, alerts and invalid fields. Element references belong to a structure snapshot, so the agent must read again after navigation or a refresh rather than reuse an old reference. The patch improves observation within this workflow; it does not replace submission checks.
Installing the release without mixing versions
The 0.7.1 release archives (opens in new tab) cover Windows x64, Linux x64 and ARM64, and Apple Silicon macOS. The documented archive workflow is to unpack the engine and Python files together. The SDK and MCP server require Python 3.10 or later; the terminal guide lists the higher requirement noted above.
The project's platform requirements (opens in new tab) specify X11 on Linux, with Xvfb acceptable, and macOS 13 or later on ARM64. CPU-default rendering does not remove Linux's display-system requirement. On Windows, libEGL.dll and libGLESv2.dll must remain beside engine.exe, as the archive instructions (opens in new tab) explain. A native --headed window requires a desktop and graphics support.
For the terminal, the documented launch command on Linux or macOS is:
./browserr-term https://browserr.dev/
For MCP, configure the client to run Python with the unpacked browserr_mcp.py, using that directory as its working directory. Call browser_start before browsing; it starts the selected engine and session.
Keep the engine tag separate from the Python package version. On November 15, PyPI listed browserr-mcp 0.7.4 (opens in new tab), but that wheel is a bootstrap installer, not engine 0.7.4. It downloads the latest GitHub engine and tooling, which was 0.7.1 at publication. BROWSERR_VERSION=v0.7.1 pins that target separately from the wheel version.
The bootstrap also has a platform-name mismatch: its macOS selection asks for a darwin-arm64 archive, while the release asset (opens in new tab) is named macos-arm64. The published code therefore predicts a missing-asset stop on a fresh macOS package installation. The manual archive workflow above avoids that naming path; a universal one-command installation claim would be premature.
The bundled updater can select a release and check its asset digest, refusing a missing or mismatched digest unless explicitly overridden. That verifies agreement with the supplied digest, not an independent guarantee of publisher authenticity. Browserr's repository carries an Apache-2.0 license (opens in new tab), while Servo 0.8.0 is listed as MPL-2.0 (opens in new tab). The repository license should not be read as relicensing its dependencies.
Where the browser still stops
The known limitations (opens in new tab) matter more than the terminal presentation for choosing a workload. Browserr documents no DRM video, WebRTC or service-worker offline support. It reports partial interactivity failures on some server-rendered React 18 hydration paths, despite support for common client-rendered applications.
More consequentially, the project reports that Cloudflare Turnstile, hCaptcha and reCAPTCHA v2 challenges do not work because their frame interactions encounter Servo compatibility problems. A human handoff cannot repair a challenge whose contents never load. A native window is useful only when the underlying page works.
The security notes (opens in new tab) also warn that Windows has no site sandbox: page code runs in the engine process. Loopback sockets are reachable by local processes by default; the driver port offers optional token authentication. Ephemeral profiles are a storage-lifecycle choice, not containment for hostile page code.
“Fail-closed” has a narrower concrete meaning than universal protection. Request interception starts disabled (opens in new tab). When enabled, it holds selected document and data requests and denies unanswered ones after five seconds; ordinary images, styles and fonts are not held. It is not a default-deny network firewall.
Pause blocks selected automation commands, not every way to change the page. Raw JavaScript evaluation and other privileged commands remain available, and page scripts continue running. The submit helper (opens in new tab) accepts a caller-supplied confirm:true flag for heuristically detected payment or CAPTCHA submissions. That flag is not independently verified human approval, nor a gate on every click, keypress or script. The calling agent must enforce authorization.
Trusted automation clients also retain host capabilities: downloads can write to supplied paths, and file-picker answers can select local files. The engine's standalone downloader connects directly rather than using Servo's proxy preferences. WebSockets bypass the configured proxy too, and DNS resolves locally except where a CONNECT tunnel delegates hostname resolution. These are operator trust boundaries, not a demonstrated website exploit.
Browserr 0.7 offers a specific integration worth evaluating: a rendered terminal view alongside named engine routing, with less ambiguous field previews in 0.7.1. The incomplete session schemas and default-driver handoff path need attention before treating it as a polished multi-browser MCP workflow. Start with the actual page and form flow an agent needs, especially its login and challenge steps. A successful terminal demo cannot establish that those steps will work.

