Buddy Bots — Grok Bot Parity Spec
Source: x.ai/bot, x.ai/news/introducing-grok-bot, docs.x.ai/grok-bot/* (read 20 Sep 2026). Grok Bot is in beta (launched 11 Aug 2026); features marked "rollout" are not available to all their users yet.
Rule: every row in "Grok Bot" must exist in Buddy Bots (parity). "Better" column is where we beat it. Phase: P1 = MVP on localhost, P2 = multi-bot + automation, P3 = apps/teams.
1. Bots
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| B1 | Bot = durable teammate: name, label, description (permanent rules), avatar | Same + structured role fields (goal, tools allowed, approval boundary) instead of one free-text description | P1 |
| B2 | Create via New / Cmd+N; auto-generate bot from a name | Same + generate from a one-line job description | P1 |
| B3 | 8 templates: Sales Outbound, Talent Scout, Paid Media, Expense Manager, Product Performance, Bug Reproduction, Account Health, Chief of Staff | Same 8 roles + template gallery | P2 |
| B4 | Edit profile, pin, hide/unhide (hidden bots keep running) | Same | P1 |
| B5 | Duplicate (copies profile, settings, skills, routines, avatar; not history/memory/attachments) | Same + optional "copy memory" | P2 |
| B6 | Share as template: public link or team-only; recipient gets a copy, never your computer/logins/history | Same + export/import as JSON file | P3 |
| B7 | Delete: removes profile, conversation, routines; shared files and logins remain | Same + optional wipe of that bot's files | P1 |
| B8 | No model picker — xAI model only | Per-bot model choice (Claude, GPT, DeepSeek, Gemini, Grok, local via Ollama) + per-bot cost cap | P1 |
| B9 | Memory: stable preferences, role context, work summaries; per-bot; corrected by telling the bot | Same + visible, editable memory page per bot | P1 |
2. Chat and collaboration
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| C1 | 1:1 conversation per bot; paste text/links/images; attach files | Same | P1 |
| C2 | / references a skill; @ mentions bots, groups, routines, connectors | Same | P1 |
| C3 | Reply in threads, emoji reactions | Same | P2 |
| C4 | Message while work is in progress; user messages pre-empt background work; "Stop now" halts (no undo) | Same + per-task cancel button | P1 |
| C5 | Group chats of 2–6 bots; editable name/members; bots decide who responds; @BotName, @everyone | Same, no hard cap of 6 | P2 |
| C5a | (not documented) | Group speaking rule: a bot may post anything a good teammate would — result, question, handoff, approval request, suggestion, warning, risk, advantage/disadvantage, disagreement, correction — provided it adds new information or moves the work forward. Blocked: echo/agreement-only, restating another bot, social filler. Jev gates each candidate message ("adds value? yes/no"); plus turn cap, claim lock, per-conversation cost budget. Warnings and risks bypass the turn cap. | P2 |
| C6 | Async bot-to-bot handoff messages (text-only) | Same + handoffs can carry files | P2 |
| C7 | Hierarchies (chief-of-staff bot managing specialists) | Same | P2 |
| C8 | Proactive pickup / follow-up on stalled work | Same | P2 |
| C9 | Search across messages, bots, groups, files, links, routines; command palette | Same | P2 |
| C10 | Attention states: "Needs attention", "Unread activity", manual read/unread | Same | P1 |
| C11 | Draft auto-save per conversation | Same | P1 |
3. Files and results
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| F1 | Up to 6 attachments; 25 MB docs/images/audio, 200 MB video | Same limits or higher (local = no upload cost) | P1 |
| F2 | Types: images, audio, video, PDF, text, Word/Excel/PowerPoint, CSV, JSON, YAML, code, HTML, email files, Jupyter | Same | P1–P2 |
| F3 | Outputs on request: documents, spreadsheets with formulas, slide decks with notes, folders of screenshots/logs, unsent draft messages | Same | P2 |
| F4 | Result cards for files, images, links, tool results; save / open source / give feedback | Same | P1 |
| F5 | Structured result: facts / assumptions / completed actions / pending approvals / open questions | Same, on by default for consequential tasks | P2 |
| F6 | Shared /workspace folder across bots | Same | P1 |
4. Computer and apps
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| K1 | One persistent cloud computer per account: browser, filesystem, terminal; keeps working when devices are off | Same (Docker container on localhost first, cloud VM later) | P1 |
| K2 | All bots share one computer — cookies, files, credentials; docs state bots are not a security boundary | Optional isolated browser profile per bot (a finance bot's logins are invisible to a sales bot) | P2 |
| K3 | Parallel bots, each with its own screen; one computer-use task per bot at a time | Same | P2 |
| K4 | "Agent Computer" live view: clicks, typing, navigation, status | Same + step timeline with screenshots | P1 |
| K5 | Takeover for password/passkey, 2FA, CAPTCHA, payment, identity check | Same | P1 |
| K6 | Secure secret input (masked, kept out of transcript) | Same + encrypted local vault | P1 |
| K7 | Persistent logged-in browser sessions | Same | P1 |
| K8 | Connectors/plugins from a Marketplace, account-wide, enable/disable | MCP-native: any MCP server works as a connector | P2 |
| K9 | Computer use for sites with no connector | Jev-first navigation (cheap/fast), vision-model fallback on BLOCKED/canvas/iframes; per-task cost and success logged | P1 |
| K10 | Maintenance: Recover from snapshot, Update, Reset | Same (volume snapshots) | P2 |
| K11 | Optional execution on the user's local computer: ask every time / always / never | Same | P3 |
| K12 | "Route egress through this desktop" (bot traffic exits from the user's own IP — their answer to anti-bot blocking) | Free on localhost; needed once cloud-hosted | P3 |
| K13 | Hardware security key (YubiKey) passthrough | Same | P3 |
5. Skills and routines
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| S1 | Skill = reusable instructions; shared private library across bots; "save this as a skill" after a task | Same; skills stored as plain markdown files (portable, versionable) | P2 |
| S2 | Teach a task: record one browser workflow (max 10 min, no audio) → draft skill to review (rollout) | Same + recorded Jev action trace replays deterministically, LLM only when the page differs | P2 |
| S3 | Routine = one bot + workflow + schedule (timezone) or event trigger | Same | P2 |
| S4 | Event triggers (Slack message, GitHub notification) with narrow match rules | Same + generic webhook + email trigger | P2 |
| S5 | Enable/pause, test run (does real work), edit, run history, delete | Same + dry-run mode (no write actions) | P2 |
| S6 | Limits: 50 routines per bot, 20 run records per routine | No cap; full history | P2 |
| S7 | Auto-pause routines after long user absence | Same | P2 |
6. Approvals, security, privacy
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| A1 | Approval card: Allow once / Deny / Always allow | Same | P1 |
| A2 | Auto-review rules: "Ask first" and "Allow automatically"; Ask first wins on conflict | Same; Jev classifies action risk in milliseconds | P1 |
| A3 | Default-gated actions: send, publish, pay, delete, permission changes, production changes, accepting legal terms | Same | P1 |
| A4 | Bot pauses and notifies instead of bypassing security checks | Same | P1 |
| A5 | Encryption in transit and at rest; training opt-out | Self-hosted option: data never leaves your machine; bring-your-own API keys | P1 |
| A6 | Cloud storage mandatory (no privacy mode) | Local-only mode | P1 |
| A7 | Full activity/action log as evidence | Same + exportable audit log | P1 |
7. Settings, notifications, platforms
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| N1 | Appearance: system/light/dark; timezone | Same | P1 |
| N2 | Per-bot notification toggle; OS and mobile push; suppressed while app is focused | Browser push first, then mobile | P2 |
| N3 | Usage display: weekly included + on-demand | Real cost per bot, per task, per model | P1 |
| N4 | Error banners, copy request ID | Same | P1 |
| N5 | Desktop apps: macOS, Windows, Linux | Responsive web app first; Electron wrapper later | P3 |
| N6 | Mobile apps iOS 18+/Android 9+: chat, voice dictation, photo capture, approvals, screen view, routines, search | PWA first; React Native later | P3 |
| N7 | Sync across devices | Same (server-side state) | P1 |
8. Teams and enterprise
| ID | Grok Bot | Buddy Bots — better | Phase |
|---|---|---|---|
| T1 | Team seats; team-only bot templates | Same | P3 |
| T2 | Admin-enforced (locked) auto-review rules; restrict local execution; managed computer setup; plugin restrictions | Same | P3 |
| T3 | DLP, proxies, network restrictions | Proxy + domain allowlist per bot | P3 |
9. Where Buddy Bots wins (summary)
- Any LLM per bot (they have no model picker).
- Jev-first browser control — large cost and speed advantage if success rate holds; measured from day one.
- Per-bot isolated logins (their bots all share one credential pool).
- Self-hosted / local-only, bring-your-own keys.
- MCP-native connectors instead of a closed marketplace.
- Transparent cost per bot and task.
- Deterministic replay of taught tasks; dry-run routines; editable memory.
- No caps on group size, routines or run history.
10. Where they will stay ahead (honest)
- Native apps on five platforms from day one.
- Subsidised usage inside a $20–30 subscription.
- Managed, always-on cloud fleet with snapshots.
- A trained-in-house computer-use model that handles canvas/iframes/desktop apps; Jev harness does not yet.
- Enterprise compliance paperwork.
11. Unknowns to verify
- Teams/enterprise docs page was not reachable; section 8 is from FAQ and security pages.
- Exact connector list in their Marketplace is not published in the docs read.
- Jev success rate on long, multi-app tasks — no published data; we log it ourselves (K9).
12. Build status — v0.3 (20 Sep 2026)
P1 and P2 are built and were tested with real keys: Jev through OpenRouter (/api/alpha/decisions, ~typesafe/jev-latest) and OpenAI models. Code: buddybots-app.zip (see its README).
Measured: SkyBench flight search 5/5 correct, ~9 s and ~$0.001 per task; 26 Jev decisions at 100% step success, median 238 ms, ~$0.00003 each. Also passed with real models: Wikipedia lookup, approval allow/deny, canvas page via vision fallback, 3-bot room with the redundancy gate, handoff, memory, PDF+XLSX reconciliation with structured report. 16 integration tests in simulator mode. Sample is small — one test site, one real site.
Closed since v0.2: B7 wipe files, C6 file handoffs, C8 proactive follow-ups, C10 mark unread, F2 PDF/Word/Excel/PowerPoint parsing, F5 structured results, K10 snapshot/recover/reset, S4 Slack + GitHub webhook events with match rules, N4 request IDs.
Added after v0.3: real desktop computer (K1/K3/K4 — X display, window manager, real Chromium window + terminal, whole-desktop live view, takeover with real mouse/keyboard; own desktop per bot when profiles are isolated), look tool (bots review their work visually), navigator hardening for live sites, access token + docker-compose for an always-on box.
Added 20 Sep (later): template library with share code / file / public link and Grok Bot template import (B6), reply streaming, audio/video transcription (F2), email trigger (S4), whole-screen desktop tool.
Still open inside P1/P2:
| ID | Gap |
|---|---|
| F3 | Office-format outputs depend on Python libraries on the host. |
| S4 | Email trigger is built but untested against a live mailbox; Slack/GitHub need the webhook URL reachable from the internet. |
| B6 | Grok template *links* import only their public fields (name, author, description); x.ai does not publish skills/routines. Full-fidelity path: the owner asks their own Grok bot to export itself (Import → My Grok bot) and pastes the reply. |
| K9 | Whole-screen clicking needs a computer-use-trained model; general models miss small targets. |
| — | One shared access token, no user accounts. Anthropic, DeepSeek, Gemini, xAI, Ollama adapters are untested. |
Tuning knobs found in testing: room speak thresholds 0.5 (owner) / 0.7 (others) / 0.65 (bot-to-bot); redundancy gate holds at ≥ 0.4; navigator stops at goal_done > 0.8; vision capped at 8 steps per run.