Feature inventory
The complete workbench
Row-Bot combines provider-aware reasoning with parent-led agents, durable document processing, goals that keep working, local memory, real tools, authenticated access from your other devices, background workflows, messaging, and reviewable extensions in one conversation-first app. Ask for a deck or a code change and the design or code folder opens beside the chat; approvals, goals, and agents stay with the conversation that owns them. This page documents what ships and where each capability fits.
The same app runs in the desktop window, in a browser on another computer, and on a phone. Conversations are where work happens: ask, and Row-Bot creates what the task needs and shows it beside the chat. Home gathers everything that is not one conversation, and nothing urgent is hidden.
App tour documentation →Home documentation →
Ask for a deck, a landing page, or a code change and Row-Bot creates the design or code folder the work needs, shown as an inline card with Open, Rename, and Undo. When a request needs a tool or account that is off, the conversation offers Turn on or Connect in place instead of sending you to Settings.
Overview opens with an Ask box, then shows what needs you with each fix in place, where you left off, today’s agents and workflow runs, and what happened since yesterday evening; approvals are answered right there. Workflows, Knowledge, Monitor, and Insights sit beside it as tabs, so background work, memory, and health are one click from any conversation.
A small card beside each conversation shows what it is working on — designs, code folders, the browser — along with its goal, its outputs, and its agents with Stop and Message. It floats beside a wide chat and opens as a sheet on smaller screens.
Press ⌘K or Ctrl+K to search conversations and their messages, commands, settings, and saved workflows at once. Plain intents in any word order, such as “connect telegram”, “turn on developer tools”, or “dark mode”, find the right action, settings switches show their state and change with Undo, and Enter runs only a result that matches what you typed.
Open a real terminal on your own shell, docked under the conversation in the desktop app, with colours, history, Tab completion, Stop with Ctrl+C, a height remembered per device, and scrollback that survives closing the dock. Open in your terminal starts your own terminal app at the code folder, and commands the assistant runs still use their approval cards.
The sidebar groups conversations by time and filters them as All, Chats, Designs, Code, or Workflows. Settings sit in six groups with search over pages and single rows, and save as you change them with Undo; appearance follows the operating system in light or dark, and any conversation exports as Markdown or PDF.
Row-Bot turns a conversation into a controlled working environment. Profiles define how an agent should operate, goals keep long-running intent visible and moving, and one authoritative parent can plan, delegate, join required results, continue through several work waves, and produce the final answer.
Profiles, goals, and agents documentation →
Give a conversation a reusable role with instructions, usage guidance, handoff contracts, workspace policy, enabled tools and skills, approval policy, and model preferences. The sidebar’s Agents dialog holds the profile library and your favourites, and profile snapshots travel with delegated work so the operating contract remains explicit.
Set a goal from the conversation details card or with /goal and Row-Bot starts at once, continuing after each turn until the work is done, with optional turn and time limits. The goal card shows the turn, time running, tokens, the latest progress reason, Pause or Resume, and Stop; a goal pauses with its reason when progress stalls, waits out provider rate limits, stops for approvals, and delivers terminal notices through configured channels.
Delegate required or background work with dependencies, duplicate-objective checks, multiple work waves, live joins, and visible capacity limits. Stop or message a delegated agent from the conversation details card or its own conversation, and answer its approvals in either place. The parent remains responsible for follow-up and the final response.
Bound long-running Agent work with durable model-and-tool round counters, repeat-stall protection, exactly-once finalization, and configurable nesting, concurrency, and active-time limits. Extra child runs wait in a fair queue instead of bypassing capacity.
An ordered inbox returns approvals, completions, retries, steering, and stop events to the parent. Checkpoints, leases, and finalization claims recover interrupted work without duplicate synthesis, delivery, or final answers.
Keep the selected conversation, including one with a design or code folder, in a compact always-available overlay on Windows and macOS. Buddy shares the conversation’s draft, model, tools, approvals, and run state instead of creating a second conversation.
Every selection stays provider-qualified from setup through runtime. Row-Bot separates model discovery, credentials, capabilities, readiness, context policy, and transport behavior so similarly named models do not become interchangeable by accident.
Models and providers documentation →
A fresh install has no preset chat, vision, image, or video model. The first run asks one question, how Row-Bot should think: on this computer, with your subscription, with an API key, or through a custom endpoint. Your pick becomes the default, and if it later becomes unavailable Row-Bot offers Reconnect or Choose another model instead of silently switching.
The composer and the Brain, Vision, Image, and Video pickers share one grouped, searchable list with billing tags, and the model pill says Unavailable with the reason when a model cannot run. Vision follows the chat model unless you choose another, and attaching an image to a model that cannot see it offers a vision model.
Discover and use Ollama plus OpenAI-compatible runtimes such as LM Studio, vLLM, llama.cpp, LocalAI, LiteLLM, SGLang, and oMLX. Local endpoints remain distinct from hosted services and can work offline when their runtime and model are local.
Connect OpenAI, Anthropic, Google AI, xAI, OpenRouter, Requesty, Atlas Cloud, MiniMax, OpenCode providers, Ollama Cloud, and custom endpoints. Capability mapping routes chat, tools, reasoning, vision, image, and video models to compatible surfaces.
Keep xAI Grok OAuth, ChatGPT/Codex, and Claude Subscription authentication separate from API-key providers. OAuth health, credential source, model visibility, and runtime readiness appear in provider settings.
Targeted, startup, and scheduled refreshes discover newly available models through provider-scoped saved credentials. Last-known-good provider rows survive failed or incomplete refreshes, while live, cached, fallback, and empty outcomes remain visible.
The composer’s context meter estimates the complete next model input, including system prompts, images, and bound tool schemas. Model-aware rolling compaction starts with safe headroom, preserves recent complete turns, and records durable status, while prompt-cache diagnostics and provider-specific reporting keep long conversations predictable.
Inspect the effective schema sent through each adapter. Gemini-compatible filtering catches untyped arrays before a request is made, while unaffected providers keep the original tool contract and explicitly requested incompatible tools fail clearly.
Row-Bot can chain built-in tools inside an agent turn, progressively discover enabled external capabilities, limit them through profiles, pause for approval right in the conversation, and return generated files to the conversation or configured delivery channel.
Progressive tools and skills →Tools and approvals documentation →Computer Use documentation →
Search the web through configured services or DuckDuckGo, read URLs, query Wikipedia and arXiv, inspect YouTube content, work with local documents, and preserve useful findings in memory or the wiki vault.
Keep core tools ready while deterministic local search finds enabled MCP, plugin, Custom Tool, and channel capabilities only when a request needs them. Auto mode binds the exact selected schema instead of unrelated catalogs; eager compatibility mode remains available without changing provider checks, profiles, approvals, budgets, or workspace policy.
Read and manage files, run approval-gated shell commands with operators explained rather than refused, and automate managed Chromium sessions under Row-Bot’s browser policy. Snapshot-bound observations, sensitive-output redaction, visible consequential navigation, takeover controls, and action receipts keep browser work reviewable.
Opt into the Computer Use beta for Windows and macOS through the pinned, checksum-verified Cua Driver 0.20.0. A Using your computer card shows the app and the latest picture, kept in memory only, with Pause to take over, Resume, and Stop, while target-window capture, semantic actions before coordinate fallback, action receipts, an exclusive interactive lease, point-of-risk approvals, blocked sensitive surfaces, and ephemeral Vision fallback constrain every session. The separately installed driver’s consent-gated third-party telemetry is disclosed before enablement and is not a Row-Bot telemetry pipeline.
Use Gmail and Google Calendar with OAuth-backed tools for search, drafting, event lookup, typed mutations, and bulk creation. Calendar operations use request-scoped clients, safe concurrent refresh, retries, ordered writes, and timeout reconciliation.
Calculate, query Wolfram Alpha, inspect weather, analyse images, search conversations, create charts, generate or edit images, and produce video through compatible provider models. Media choices remain scoped to the correct surface, with xAI image options discovered from live model capability metadata.
Useful recall needs more than an opaque summary. Row-Bot stores structured local records, combines semantic, lexical, and graph signals, and keeps review state and trace information available when you need to understand why something returned.
Knowledge and memory documentation →
Represent people, projects, preferences, facts, events, documents, and other entities as typed nodes and relationships. Explore them on Home’s Knowledge tab, where a WebGL graph shows up to 2,000 memories (5,000 with Show all), focuses a memory’s neighbourhood, and filters by type, beside the review queue, bulk actions, Add memory, and activity logs. Vague relation types are refused before saving, and inconsistent records can be repaired.
Combine configurable embeddings, FAISS semantic search, FTS5 lexical matching, and bounded graph expansion. Local embedding startup is cache-only, downloads and repair are explicit, and missing models fall back quickly to lexical and graph retrieval instead of stalling the request.
Stream bounded uploads into a persistent SQLite queue, deduplicate content, make the complete batch searchable before extraction, and publish sharded vector manifests atomically. Pause, resume, cancel, retry, repair, and restart without losing safe progress.
Extract knowledge from PDFs, spreadsheets, JSON, TSV, images, and common text formats through bounded map-and-reduce checkpoints. Source metadata stays attached and final document, graph, and Wiki Vault updates commit idempotently.
Export an Obsidian-compatible local wiki with readable article names, synchronise selected knowledge, and let the idle-aware Dream Cycle consolidate memory or prepare bounded improvement proposals while preserving busy-state and review controls.
Turn repeated tasks into scheduled or manually triggered pipelines without turning them into invisible automation. Profiles, delivery choices, conditions, approvals, run history, and recovery state stay attached to the workflow.
Choose a profile, model, tool policy, prompt, schedule, persistent conversation, conditions, approval mode, and where results go. The builder keeps steps in an ordered list with the schedule written in words, suggests variables and step outputs as you type {{, and keeps the graph editor one toggle away; webhook triggers show their address with Copy and accept their secret as a header.
Pause consequential actions for approval in the app or supported channels. Results always arrive in the app, plus any configured channels you tick; workflow-level delivery defaults distinguish no destinations from inherited destinations, and channel retries retain terminal notices until they can be reconciled.
Follow a run live in its drawer, run a failed workflow again from Overview, and inspect scheduled runs, subtasks, pipeline state, errors, completion summaries, and recovery diagnostics on Home rather than relying on ephemeral notifications.
Stop connects the active generation to provider responses, subprocesses, tool waits, browser work, MCP calls, voice turns, and child agents spawned by that generation. Unrelated runs remain untouched.
The desktop app, a browser on another computer, and a phone open the same app against the same Row-Bot host. The host remains the system of record, a small screen changes layout rather than authority, and every signed-in device can be reviewed and signed out.
Remote access documentation →Phone layout documentation →
Connect a phone or computer from Settings › Devices & remote access. Row-Bot detects the ways in — Tailscale for private HTTPS, the same Wi-Fi, or the internet through your public link — then shows one QR code backed by a one-time invitation that renews itself, waits for the device, and shows how to install Row-Bot on that phone.
Small screens get one header, the sidebar as a drawer with the same filters, panels as full-height sheets, and every Settings page. Create conversations, attach files, choose profiles and models, approve work, and stop active runs away from the desk.
See each signed-in device by name, which one you are using, and where and when it was last seen in plain words. Rename a device or sign it out, which ends every session of that device; invited devices renew their session in the app and never revive a revoked one.
Create origin-bound, expiring, one-time invitations from the app or CLI. Instance-isolated HttpOnly sessions renew authoritatively, record secret-free audit events, and support logout plus immediate session or device revocation.
Run row-bot serve for a long-lived headless host. HTTP requests and the live event stream share exact-origin, allowed-host, trusted-proxy, and authenticated-owner gates; server loopback does not bypass authentication.
Save exact owner-controlled HTTP or HTTPS browser addresses so admission, invitations, and later removal take effect without a restart. Environment-managed origins remain read-only, assigned non-private interfaces carry explicit exposure warnings, and Row-Bot cleans up only the Tailscale routes and tunnels it owns.
Messaging adapters share one delivery engine while preserving the limits and native interaction model of each platform. Channel work uses the same conversations, profiles, goals, approvals, media handling, and final records as desktop turns.
Channels and voice documentation →
Set up Telegram, Slack, Discord, SMS, WhatsApp, Google, X, GitHub, and plugin sign-ins through numbered steps in one sheet, with an explicit test message to yourself and a live WhatsApp QR code. A channel that needs a public address opens it only when you configure that channel and shows where it is reachable.
Coalesce edits, keep typing alive, bound previews, split long output safely, respect rate limits, clean up previews, and send a fresh final message when an edit cannot complete. Discord, Slack, Telegram, and WhatsApp use the best supported path.
Approve or deny through Discord buttons, Slack Block Kit actions, Telegram inline controls, WhatsApp resume flows, or explicit SMS YES/NO handling. SMS intentionally remains final-text-only with safe chunking.
Use local Whisper speech-to-text and Kokoro text-to-speech for dictation and read aloud, or opt into compatible realtime voice providers to talk. Authenticated browsers can capture, transcribe, and play speech locally to that session without starting the host microphone service or retaining browser media.
Receive images and documents, return generated files, show reactions and health state, and deliver compact once-only goal or child-agent completion notices. Failed terminal notices remain available for retry.
Design and code work happens in the conversation that asks for it. A design or code folder belongs to that conversation and opens in its Design or Developer panel beside the chat, so you can ask for a change, watch it land, review it, and keep going. Generated designs stay editable; repository changes stay reviewable.
Designs documentation →Code folders documentation →
Make decks, documents, landing pages, app mockups, and storyboards from a request, Add resource, or Home’s New design. The Size menu offers 16:9, 4:3, square, A4, and phone and re-fits every page, anything unusual is one request away, and brand colours and fonts can come from a website.
Switch between Preview and Edit, select any element to ask Row-Bot for a bounded change, or double-click text to edit it in place. Pages, undo and redo, saved versions, and a review that fixes issues one by one or every safe one at once keep iteration reviewable.
Present full screen, publish a link with Copy and a locally drawn QR code, send to a channel, or prepare an X post. Export to PDF, PNG, PowerPoint, or HTML in one click into the workspace’s Exports folder, while the sandboxed runtime constrains generated scripts and critique and repair loops check the work before handoff.
Generate or edit images, add provider-compatible video, insert charts, attach references, and keep assets with the design before the first draft or a later revision.
Ask Row-Bot to use a folder you already work in or to clone a repository, and it binds the folder once you choose where; Add resource can also start a draft, register an existing folder, or clone. Worktrees keep a conversation or delegated run separate from unrelated work.
Review diffs and the agent’s change sets in Changes, browse Files, run detected checks and reviewed commands in Run, and branch, commit with a suggested message, push, and open pull requests in Git. Install, network, delete, commit, push, and PR actions require approval.
Run commands and edits in a shadow copy, inspect the result, and import approved changes back into the real repository. Import gates protect the workspace from unexpected sandbox output, and sandbox network access stays off, asks, or on by your choice.
Profiles, goals, child agents, tool policy, write locks, cancellation, and approval reasons work the same in a conversation with a code folder, so long-running implementation follows the same operational model as other Row-Bot work.
Skills, plugins, MCP servers, and Custom Tools add capability through separate contracts. Enablement, permissions, configuration, health, and profile policy determine what becomes available to an agent.
Discover public skills, install, enable, and pin reusable instruction packages, invoke slash commands, and let Row-Bot select up to five relevant enabled skills per parent task or child Agent. Reopened tasks restore their skill state, while skills remain instructions and cannot grant tools or bypass profile, approval, budget, or workspace boundaries.
Install reviewable plugins that provide native tools, packaged MCP servers, bundled skills, or channels. Manifests declare permissions, settings, secrets, authentication, health checks, versions, and supported extension surfaces, and their disclosures appear before installation.
Connect stdio, HTTP, or SSE servers with capability metadata, safety classification, requirements, probes, conflict handling, and approval gates. Add and connect tests a server, lets you accept its tools, and enables it in one pass, while marketplace metadata, including catalog entries such as Xquik, remains intact through configuration edits.
Turn a repository, local folder, or current code folder into a reusable tool through Source, Inspect, Test, and Enable stages, then manage it in Settings › Tools › Custom tools. Its commands use the standard approval card, and safer Git setup, isolated environments, command checks, and removal keep experiments bounded.
Local-first means durable application state belongs to the user, external routes are explicit, and consequential actions remain reviewable. It does not pretend that a hosted model or online tool can be used without sending that service the data required for the request.
Privacy and safety documentation →
Store conversations, profiles, goals, orchestration records, workflows, document queues and indexes, memory, designs, plugin state, code context, and owner access records under the Row-Bot data directory.
Use the operating-system credential store on desktop or an isolated encryption key with encrypted provider and channel records in server deployments. Read-only secret directories, safe fingerprints, and redacted status keep plaintext credentials out of application metadata and diagnostics.
Classify destructive, networked, private-data, metered, installation, Git, shell, browser, channel, and MCP actions conservatively. Approve or Deny in place — in the conversation, the attention indicator’s list, Home’s Overview, a floating notice, Buddy, or a supported channel — with bounded reasons. Approvals wait until answered and survive restarts, while the original action payload is retained for the decision.
Row-Bot has no required account, hosted sync service, or first-party telemetry pipeline. Optional providers, channels, tunnels, updates, search, and external tools remain individually visible choices. The opt-in Cua Driver has separately disclosed, consent-gated third-party telemetry; it is the reviewed exception, not Row-Bot telemetry.
Back up now writes one zip to the workspace’s Backups folder without keys, sign-ins, sessions, caches, logs, or runtimes, and lists what to set up again. Restore from backup checks the archive, applies it at the next start, keeps the current profile aside, and rolls back if anything fails.
Select conversations in the library with filter-aware bulk controls and remove them through one cleanup path. Active work is cancelled before deletion, while dirty Developer recovery state, code folders, and design ownership are preserved.
Windows and macOS are first-class desktop targets, Linux provides the browser/server application through a verified XDG user install, and the official multi-architecture container provides a hardened headless path for private servers and VPS hosts.
Installation documentation →Docker and VPS operations →Troubleshooting documentation →
Use the Windows x64 installer, macOS Apple-silicon DMG, or verified Linux one-line installer and tarball. Published artifacts, checksums, and platform trust details are listed with each release so signing or notarization is never implied.
Run the locked non-root amd64 or arm64 image with a read-only root filesystem, dropped capabilities, bounded logs, a persistent data volume, and a separate persistent secrets volume. Compose, VPS, Caddy, systemd, backup, upgrade, and rollback examples keep operations explicit.
The first run asks one question — how should Row-Bot think? — and opens Home once you choose a model. Setup Center then resumes everything else across providers, credentials, tools, optional integrations, and runtime checks, distinguishing missing configuration, missing dependencies, unhealthy credentials, and unsupported models.
Home’s Monitor shows health tiles for the model runtime, channels, MCP and tools, the scheduler, knowledge, and the system, with each problem’s fix in place: restart a channel, renew a sign-in, choose a default model, or open the exact setting. Insights suggest one action each, and the attention indicator appears only when something needs you.
Check signed release metadata, download compatible updates, validate installer contents, and preserve local data through upgrades. Deeper installer verification covers clean installs, repair, upgrade, uninstall, and installed-package smoke checks.
Get Row-Bot
Start locally, add providers and integrations deliberately, and let the workbench grow with the work.