ToolBraid Companion.

• Project access profiles: allow exact origins, choose read-only or read/write access and exclude domains.
• Reviewable bug reports: explicitly capture new console errors and failed request metadata; add notes and an optional screenshot, review the draft and download JSON.
• Native WebMCP scenarios: save exact inputs and assertions, review the schema, approve a run and compare results.
• Incremental context: request changes after a full page baseline, with a safe full-read fallback.
• Connection checklist: inspect observed extension, page and companion state. AI account access remains a separate check.

Profiles restrict browser operations, not companion files or desktop tools. Bug capture uses the debugger; screenshots are unredacted and nothing is automatically shared. WebMCP scenarios require native API support. Usage and limits.

• Read and act: inspect visible text, links, forms and controls; click, type, fill, select, submit and scroll. Recheck page state around actions.
• Manage tabs: list, open, activate, navigate, go back or forward, reload and close an explicitly selected tab. Wait for a load or an expected URL.
• Work across frames: list accessible frames, select one and return to the top frame, within granted site access.
• Use native WebMCP: discover tools and their schemas registered by a compatible page or frame, then execute an exact page-bound tool call.
• Handle transfers: attach an explicitly granted local file; list, wait for, cancel or reveal browser downloads. A completed download can become a one-use upload grant.
• Recover deliberately: wait for a stable page or an expected control; detect rate limits and human-verification screens and hand control back when needed.

Native WebMCP needs both a supported browser API and a compatible site. Ordinary page tools work independently. Site permissions still apply; ToolBraid does not bypass CAPTCHA, login or MFA.

• Inspect desktop apps: list visible, enabled windows and their exposed Windows UI Automation controls.
• Operate exact controls: invoke a revalidated control or set the value of a non-sensitive field.
• Grant selected files: use an owner file picker to issue short-lived, one-use handles; revoke an unused grant.
• Grant a folder: list, read, create, move and archive files inside the approved root, then revoke access. New writes and moves do not overwrite existing files.

Desktop access depends on the app exposing accessibility controls; this is not unrestricted remote-desktop control. Folder listings are bounded to 3 levels and 100 entries; reads support files up to 1 MiB in chunks up to 256 KiB, and new writes up to 512 KiB.

• Inspect rendered content: capture page screenshots and supported rendered images, video frames or audio for analysis.
• Understand custom UI: use a separately configured multimodal provider to inspect visible canvas content. Coordinate actions are bound to a fresh screenshot, viewport and page.
• Use accessibility handles: inspect the browser's actionable accessibility tree and activate an exact, freshly revalidated node.
• Run local media jobs: trim, concatenate or transcode media, burn subtitles and generate thumbnails. Read job progress or cancel a job.
• Choose output formats: the job schema supports H.264, H.265, VP9 and AV1, with MP4, MOV, MKV and WebM containers and PNG, JPEG or WebP image output where applicable.

Advanced browser operations use the debugger permission. Visual analysis needs a compatible provider and its own configuration or credentials. Media editing requires FFmpeg installed separately; available codecs depend on that build. Neither FFmpeg nor an AI model is bundled. Inputs and outputs use file grants, not arbitrary shell commands.

• Research across tabs: create a bounded local workspace, attach selected tabs and capture page snapshots with source, time and evidence fingerprints.
• Build citation bundles: retain provenance, deduplicate evidence and retrieve it in pages for your assistant's answer.
• Keep useful context: store, query, export or forget bounded local semantic records with citations and freshness information.
• Remove retained research: explicitly forget workspaces and records when they are no longer useful.

Memory search is deterministic and lexical, not model training or a vector database. Research covers the tabs you attach, not a complete website crawl. Results sent to a cloud AI may leave your PC.

• Build durable missions: define up to 64 exact tool steps with dependencies, run or resume from checkpoints, inspect state, cancel and clear completed missions.
• Keep reusable demonstrations: store, list and forget bounded sequences of semantic steps. Turn an example into a versioned declarative adapter, review it and explicitly enable or disable it.
• Preview without acting: shadow replay compares recorded observations without executing the workflow or mutating the page.
• Manage scheduled calls: add one-shot or recurring exact tool calls, list them, pause, resume or delete them, and run due calls through the schedule tick tool.

Scheduled calls need an active runner/client to trigger due execution and a connected extension for browser work; this is not a hosted, always-on service. Saved workflows do not guarantee completion, and interrupted mutations must not be blindly replayed.

• Keep up on X: inspect rendered replies and mentions, identify what needs attention and draft or send the response you request—while you stay focused on your project.
• Work with X composers: tools cover posts, replies, likes, reposts, quotes, image/video attachments, media controls and article draft editing.
• Monitor quietly: opt into one X conversation or notifications tab, keep a local inbox and optionally receive count-only notifications. No model calls or automatic replies are made by this monitor.
• Use other site-aware adapters: the package also includes adapters for Google, YouTube, Reddit, Discord, WhatsApp, Telegram, Instagram and TikTok, alongside general page tools.

Adapters work with the rendered site, not official service APIs. Actions depend on the current view, account eligibility, permissions and site changes. X reading and authorized replies have live validation; media and article workflows have fixture coverage, not complete live certification. Quiet monitoring needs the browser open and is not a DM reader or account archive.

• Platform: Windows x64 companion and a Manifest V3 extension for Chrome or Edge. The extension declares Chromium 120 as its minimum; native WebMCP needs additional browser support.
• Connection: a local stdio MCP server with authenticated native messaging to the browser. The complete Windows ZIP includes its Node.js runtime.
• AI choice: use Codex or another local MCP client, optional ChatGPT panel chat through Codex, or a tool-capable local model in an external host. See connection options.
• Access controls: explicitly enable AI control and site access, grant local resources, and pause new browser commands. The built-in chat requests approval for browser changes; external clients use direct-control access and their own approval policies.
• Data handling: results go to your chosen AI client and any configured media provider. Local storage does not mean all processing stays local. Read the privacy policy.

ToolBraid includes no model subscription and does not transfer subscriptions between providers. Microsoft signs the Store companion; the alternative manual ZIP launcher is unsigned. The number of available page tools changes with the connected page, permissions and session.