Open Crabs 🦀

v0.3.37 JUST DROPPED 🦀🔥 OpenCrabs and Xiaomi Mimo have done a partnership to give you all FREE tokens 🤝. -> 2 weeks completely FREE(up till June 27th 00:00 UTC). No API key. No credit card. Just install binary and go. -> 5 models. mimo-v2.5-pro with 1M context. Thread 🧵👇

@MercyDeGreat Open-souce is the way.

Me explaining to people that Moonshot released an AI model kimi k3 that mogged fable 5 of Anthropic and GPT 5.6 of OpenAI in benchmark Open source AI is currently winning

@declutterKE @the_markimooo @theo @CommandCodeAI Thank you @declutterKE. Kimi k3 has been a perfect match to us such an amazing and intelligent model.

@the_markimooo @theo try @CommandCodeAI for its tool repair autohealing or something that selfevolves while still allowing for finegrained vibecoding like @opencrabs

Kimi k3 is an incredible model. It is not an incredible value. In most tasks, it comes out to roughly the same cost as GPT-5.6 Sol. K3 is half the price of 5.6 Sol per token. GPT-5.6 uses half as many tokens. Price evens out. GPT-5.6 is 2x faster TPS, so it gets work done ~4x faster than K3 at roughly the same price. I still love K3 and will be using it for a TON of stuff. I'm just tired of people pretending it's way cheaper when it's not.

Our system is working pretty smoothly with Kimi k3 as custom provider or the new Moonshot AI built-in provider @Kimi_Moonshot launched on the last release: x.com/opencrabs/stat… Its robust and consistent, sometimes we ran into rate limits but our system also handle that well making it work on 3rd or 4th retry.

@theo I feel like K3 is being limited by the harness I’m using. Currently I’m using opencod and it feels a bit clunky where it randomly stops working for no reason (unsure if that’s the model or the harness) it also misses a lot things bc of lack of computer/browser use for me

🦀 Get v0.3.69 Single Rust binary recommended: 🐙 github.com/adolfousier/op… cargo install opencrabs 📦 crates.io/crates/opencra… Built-in Moonshot AI provider, tap-to-send follow-up suggestions across every channel, MiniMax-M3, and click-to-expand. Grab it at opencrabs.com

v0.3.69 JUST DROPPED 🦀🔥 🌙 Built-in Moonshot AI (Kimi) provider 🔀 API plan + Coding plan endpoint picker 💬 Tap-to-send follow-up suggestions on TUI, Telegram, Discord, Slack, WhatsApp 🧠 MiniMax-M3 wired in across the stack 🖱️ Click a tool or reasoning block to expand it 💰 Kimi K3 now prices correctly on /usage 🖥️ Command Code CLI provider 19 COMMITS • 4 FIXES • 12 FEATURES • 5,018 tests Get it 🧵👇

OpenCrabs v0.3.68 is out. This release is all about making plan mode actually usable in practice. I have been iterating on the plan lifecycle engine from v0.3.67 helping to make the new plan mode @leshchenko1979 contributed too, more robust, and this is where it starts feeling like a real product feature instead of a technical beta. The big one is the persistent plan card. Before, every turn would spawn a new checklist message, cluttering the chat. Now there's one card that updates in place as tasks progress. You approve the plan, it shows the checklist, tasks get checked off, and the card stays at the bottom of the chat. No more scrolling through five different checklist messages to see where you are. Then there's OpenCrabs self-approval. If you tell the crabs "go ahead, you can approve your own plan", it now actually does that. It grants itself autonomy and can self-approve without waiting for you to click Approve. This is huge for long-running tasks where you want the agent to just execute without babysitting. The /plan command is a small quality-of-life thing but it matters. Instead of typing /plan, hitting enter, then typing your query in the next message, you can just do /plan build me a REST API and it enters plan mode with that as the intent. One step instead of two. The rest is 16 fixes to the plan flow. Things like making sure that, the approval button on Telegram works, the tool loop actually starts when you click Approve, streaming the plan execution to Telegram so you see progress, hiding unfilled fields in the plan scaffold (design vs plan) so it doesn't look broken, and fixing the hashline edit tool to use 4-char hashes instead of the old format. 21 commits. 3 features. 16 fixes. 4,964 tests. Grab it at opencrabs.com

This is actually pretty dope! @opencrabs is incredible effective on @trello too. Adding or updating tasks and even tracking comments.

OpenCrabs is the stack here, check it out.

2 недели внедряю miidas.ru на @opencrabs - ИИ-финансист для малого бизнеса. Три сценария: (1) клиент не может описать свой процесс (2) не знает возможностей ИИ - для чего он подходит, а для чего - нет (3) забывает про бота через два дня #automation #smb

@MercyDeGreat LFG sir

I am back to your TL again Reply if you can see this

New plan mode on @opencrabs - still polishing, but looks promising!

@opencrabs 🔥🔥🔥

v0.3.67 JUST DROPPED 🦀🔥 Huge shout out to @leshchenko1979 and @syafriadit1 for the great contributions. 🧠 Full plan lifecycle engine with design and checklist tracks 📋 /plan, /show-plan, /execute, /discard on TUI and Telegram 🌍 Multilanguage prompt analyzer with 6 language packs 🔄 Web search rotates User-Agent and retries DuckDuckGo 403s 💬 Telegram retries on 429 rate limits 📊 Flow chrome shows plan title, checklist progress, and goals 🔧 Config hot-reload now surfaces recovery to the user 🎯 Channel context headers include chat_id and thread_id 🗑️ Dropped orphaned task_manager and SQLite plan path 52 COMMITS • 20 FIXES • 20 FEATURES • 4,938 tests Get it 🧵👇

🦀 Get v0.3.68 Single Rust binary recommended: 🐙 github.com/adolfousier/op… cargo install opencrabs 📦 crates.io/crates/opencra… Plan mode UX lands: persistent plan card, agent self-approval, /plan <query>, and 16 fixes to the plan flow. Grab it at opencrabs.com

What's in v0.3.68: • Agent self-approval: when user grants autonomy, agent can self-approve without waiting • Persistent plan card: single card that updates instead of new checklist per turn • /plan : enters plan mode with the query as the intent, no extra steps • Rich report intermediate: deliver rich reports instead of folding them • 429 retry: retry rich flow edits on rate limits instead of falling back to HTML • Stream plan execution: plan Approve-button execution streams to Telegram • No duplicate presentation: stop duplicating plan presentation and approval • Run tool loop on Approve: tool loop actually starts when you click Approve • Approve on tasks: checklist-mode plans approve on tasks, not design .md • Hashline edit: widen hash to 4 chars, strip real read prefix 21 COMMITS • 16 FIXES • 3 FEATURES • 4,964 tests

v0.3.68 JUST DROPPED 🦀🔥 🧠 Agent self-approval when you grant autonomy 📋 Persistent plan card instead of per-turn checklists 🎯 /plan enters plan mode with intent 🔄 Rich report intermediate delivery 🔁 Retry on 429 rate limits 📊 Stream plan execution to Telegram 🎨 Hide unfilled plan-scaffold fields ⌨️ Plan keyboard only at turn end 🔧 16 fixes across plan flow and Telegram 21 COMMITS • 16 FIXES • 3 FEATURES • 4,964 tests Get it 🧵👇

🦀 Get v0.3.67 Existing users: Run /evolve to update. Single Rust binary recommended: 🐙 github.com/adolfousier/op… cargo install opencrabs 📦 crates.io/crates/opencra… Plan lifecycle engine, multilanguage analyzer, flow chrome with plan title and checklist progress, web search retry, Telegram rate-limit handling. Grab it at opencrabs.com

What's in v0.3.67: • Plan lifecycle engine: plans live as NoPlan/Editing/Active with a tool-loop gate that enforces write policy, a durable pre_init_editing sidecar, and archive-on-last-complete • Plan tool contract: mode (design or checklist), add_tasks, start/complete with deterministic Editing refusals, and acceptance criteria that set the session goal • Telegram flow chrome: F1 uncollapsed shell, F2 always-visible title with full checklist, F3 plan-state copy lock, goal chrome, and per-heading plan prose • Approve/Discard flow: inline buttons on Telegram, /show-plan overlay on TUI, Approve validator scans the session .md for Context labels and numbered steps • Dual-track prompts: design-track soft-nudge, state-branched compaction recovery, and the shared SESSION PLAN vs CHECKLIST track-selection table • Multilanguage analyzer: shared PromptAnalyzer with soft-nudge on TUI and Telegram, 6 language packs (EN/ES/FR/ID/PT/RU) • Web search retry: rotates User-Agent and retries DuckDuckGo 403s • Telegram rate-limit retry: retries telegram_send on 429 with backoff • Merged flow message: pre-flow status bubble gone, thinking and Working-on previews ride the flow header from the first activity tick 52 COMMITS • 20 FIXES • 20 FEATURES • 4,938 tests

Amigo penso que ias adorar a @opencrabs, por favor testa. Imagina isso mas local no teu pc, telegram, WhatsApp, Slack ou Discord. Tu personalizas a tua maneira mas já de início devia surpreender te. Não é uma LLM mas tem uma memória brutal que personalizas no soul.md e mantém se com qualquer modelo que usares.