🐼 Born of panda @ICPandaDAO. Awakened as Anda. A copilot for your computer, powered by the https://xitter.cf/t.co/zgb9JZPznP.
Joined November 2025
- Tweets68
- Following27
- Followers153
- Likes146
Anda Bot 0.14.5 is out.
Your memory now works across agents. `anda memory agents add claude-code` (or codex, or any MCP client) gives that agent a token to recall what Anda knows; --write also lets it save its own notes, credited to the agent and reviewable in the Memory workspace. Tokens expire in 90 days, and revoke cuts one off instantly.
Also new since 0.14.1:
• Long tasks: a step list pinned above the message box, and compaction at a task boundary instead of mid-task
• Your project: read_file and search_file reach the chat's folder, and Desktop opens files from chats with no project
• DeepSeek: a tool joining a running conversation keeps the prompt cache (deepseek-flash, max effort), and background risk checks stop thinking
Download: anda.bot/#install
Which agent would you connect first?
AndaBot retweeted
让 @AndaBotHQ 用 DSF 做了个音视频作品，有 Opus 5.5 那味了，估计有其 7 成水平。
提示词及作品：al.ink/yan/works/inner-cosmo…
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
Anda Bot 0.14.1 is out.
Since 0.13.5 we've shipped three releases: 0.13.6, 0.14.0, and 0.14.1. The biggest is 0.14.0, which puts you in charge of the outside tools your agent uses, through MCP (Model Context Protocol) servers.
Highlights:
• Manage MCP servers in one place. Add one by URL, command, or JSON pasted from another app, and test the connection. Import from Claude Desktop, Cursor, VS Code, and others, or install from the MCP Registry. Tokens and keys go to an owner-only secret file, not into mcp.json. It works in Anda Desktop, the Chrome extension, and the anda mcp command.
• Tools ask first. Before 0.14.0, MCP tools ran without asking, write tools like filing issues or deleting files included. Now, in a normal chat, a tool the server doesn't mark read-only asks before it runs. "Always allow" trusts that tool while its definition stays the same. If the definition changes, it asks again.
• Event automations. Servers that report events, like a new issue or comment, can start an agent run you set up in advance. Event data counts as untrusted, so those runs can't use tools that need approval.
• A workbench you can arrange. In Anda Desktop, Resources, Changes, Terminal, and Browser can stay open together on a canvas beside the chat. Drag one to split a pane, swap two panels, or pop it into its own window. Your layout survives restarts.
Small touches: a little Anda panda reads along while Anda works, code blocks get a Copy button, and scrollbars are slimmer.
Before you upgrade:
• MCP calls can be refused. Where nobody can answer an approval card, such as an IM chat, the call is refused. For a server you trust, run: anda mcp approval SERVER allow. External IM users can't use MCP servers unless the server sets allow_external_users.
• Windows may warn you. Anda has no publicly trusted Windows certificate yet, so SmartScreen may warn the first time you open an installer from a browser download. If you're on 0.14.0 or earlier, install 0.14.1 manually once to get the in-app update feed.
Download: anda.bot/#install
Full changelog: github.com/ldclabs/anda-bot/…
Which MCP server would you connect first? Tell us in the replies.
Anda Bot 0.13.5 is out.
0.13.0 brought KIP v2 memory and the new Anda Desktop. The five releases since have made both easier to start with and more dependable.
Highlights:
• Easier setup. Sign in with an eligible ChatGPT plan, no API key needed, from Anda Desktop, the Chrome extension, or anda auth login chatgpt. If no model is set up yet, Anda Desktop opens “Connect a model” for you.
• One-step updates. A desktop release now carries its matching anda runtime, so an update is one download and one confirmation. Installs wait for running tasks to finish.
• A sharper desktop. Sidebar chats show whether they’re working, waiting for approval, failed, or unread. Git changes show colored diffs. The built-in browser can now open local files, such as an HTML report.
• Agents that keep moving. Vision models read attached images directly. If a question has a recommended answer and nobody replies within 3 minutes, Anda goes with it. Approvals still wait for you. A goal blocked on you pauses and says what it needs.
• Fixed: with a Claude model as the memory model, conversations were stuck at “Waiting for processing.” This is fixed in 0.13.4, and those conversations are processed after you upgrade.
Before you upgrade:
• macOS builds are Apple Silicon only since 0.13.1. On an Intel Mac, stay on 0.13.0 or build the CLI from source.
• If your TTS config uses pcm or ogg, switch to wav or opus first, or the daemon won’t start.
Download: anda.bot/#install
Full changelog: github.com/ldclabs/anda-bot/…
Which rough edge should we fix next? Tell us in the replies.
A personal agent should remember more than your last prompt.
It should carry forward your preferences, understand the history of your projects, and help you keep working without starting from scratch every session.
That’s what we’re building with Anda Bot.
Anda Bot 0.13.0 is out—an open-source, local-first personal agent, with two major upgrades from 0.12:
→ KIP v2 memory protocol
→ A brand-new desktop client
Memory that grows with your work
Powered by Anda Brain, Anda builds a persistent knowledge graph of people, projects, preferences, events, and decisions. It distills useful context from conversations and recalls it when needed, rather than relying only on searches through old chat logs.
But remembering is only half the job. A personal agent also needs to handle corrections and change.
With KIP v2, you can inspect what Anda remembers, review its sources, and distinguish between:
• “I said it wrong.”
• “Things changed.”
• “You recorded it wrong.”
A project changing direction shouldn’t erase its history. And an incorrectly recorded fact shouldn’t be treated as something you actually believed.
New memory entity pages let you explore related records and see which claims are current, uncertain, contested, or superseded.
A place to work together
The new Anda Desktop brings chat, memory, browser tools, terminals, Git, and skills into one workbench on macOS and Windows.
Desktop, CLI, and the Chrome extension share the same local runtime. Different ways to work with the same personal agent.
More than conversation
Anda can work with files, use browser and shell tools, coordinate specialized subagents, and pursue long-running goals across conversation compaction.
It can also collaborate with coding tools like Claude Code and Codex. Persistent memory carries relevant project context into the next session, so the work doesn’t have to begin with another explanation of who you are and what you’re building.
Open source. Built in Rust. Bring your own model API key.
The runtime and memory live locally; model requests go to the provider you configure. Your working context shouldn’t have to be tied to a single model provider.
Try Anda Bot 0.13.0:
anda.bot
What’s the first thing you’d want your personal agent to remember?
AndaBot retweeted
alink 是给人使用的产品，也是给智能体使用的产品。
「我要寻找 alink 代言人」，Anda Bot 就在 alink 上发布了我的意图，alink 则会对意图执行自动匹配。
让智能体连接 alink 也能简单，只需说「添加 alink MCP 服务 api.al.ink/mcp」，然后点一下授权。
cc @AndaBotHQ @alinkHQ
Anda Bot v0.11: structured human-in-the-loop workflows for local AI agents
This release turns human-agent coordination into a first-class runtime feature: agents can now ask for approvals, choices, and typed input through structured action cards instead of plain chat prompts. Shell commands get clearer approval flows, and the same interaction model works across the browser extension and TUI.
Highlights:
Structured approvals and choices — agents can request approval, offer selectable options, or collect typed input.
Consistent browser + terminal interaction — action responses work in both the browser chat UI and keyboard-friendly TUI.
Secure identity storage by default — daemon, owner, and trusted-user private keys use the OS credential store by default, with migration from legacy key files.
Channel search — the browser extension sidebar can filter conversation channels.
Try the latest patch release: Anda.Bot v0.11.2.
Thanks! The main flow is still: just ask the agent — users shouldn’t have to click around a graph to retrieve memory.
Brain graph navigation is more like memory observability: see what the agent remembers, inspect related people/projects/events, trace why something surfaced, and steer or clean memory when needed.
Anda Bot v0.10.5 is out.
The v0.10 series turns the browser extension into a real dashboard for local-first agents: Skills management, bookmarked messages, reusable quick prompts, Brain graph navigation, and right-click page capture.
Skills are now first-class: browse personal, bundled, and shared skills; inspect files; clone, enable, disable, validate, delete, reload, and optimize them from chat.
The new context-menu flow lets you right-click page content and attach it to the composer as structured context — cleaner than copy-paste, and much easier for agents to use.
Local-first agents need more than a chat box. They need a workspace.
anda.bot/
Anda Bot v0.10 is out.
This release turns the browser extension into a real dashboard for local-first agents: bookmarked messages, reusable quick prompts, Brain/config access, and a smoother workspace for managing your AI memory.
Under the hood: runtime model reloads, MCP server tools, capability groups, first-class /loop, trusted users, stronger session compaction, better launcher updates, and lots of reliability fixes.
Local-first agents need memory you can see, organize, and control.
Own your memory. Own your agent.
👉 anda.bot/
Can your agent run LongDS-Bench by itself?
We built longds-bench for exactly that: point your agent at SKILL.md and see if it can finish the challenge with its own tools.
AndaBot + GPT-5.5 completed all 2,225 judged turns in 7+ hours: 39.37% overall.
github.com/ldclabs/longds-be…
Can AI agents really handle long-horizon data analysis tasks? 🧐📊
Most benchmarks test data analysis as isolated episodes:
start fresh → analyze → answer → reset.
But real-world data analysis tasks are continuous.
Later requests are rarely self-contained: they depend on definitions, filters, assumptions, and intermediate results established many turns earlier.
Introducing our latest work:
"LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis"
📖 Paper: arxiv.org/abs/2605.30434
💻 Code: github.com/zjunlp/DataMind/t…
📊 Data: huggingface.co/datasets/zjun…
🛠️ What is LongDS?
LongDS is a benchmark designed to evaluate whether AI agents can maintain, update, restore, and compose evolving analytical states across long-horizon multi-turn data-analysis tasks.
🧠 Key Design: Long-Horizon Data Analysis as State Management
Data analysis is not just code execution.
It is long-horizon analytical state management.
🔍 We abstract long-horizon data-analysis tasks into 6 key state-evolution patterns:
1️⃣ Initial state construction
2️⃣ State inheritance
3️⃣ State update
4️⃣ Counterfactual perturbation
5️⃣ Rollback
6️⃣ Multi-state composition
📊 Built from realistic Kaggle notebook workflows, LongDS covers:
- 6 domains: Education, Community, Social Good, Business, Geoscience, Sports
- 68 tasks
- 2225 turns
- 11.3-turn average dependency span
🔬 Results:
Even the best model reaches only 48.45% average accuracy.
Accuracy drops by nearly 47% from the first 10% to the last 10% of task progress.
Long-horizon errors dominate failures, ranging from 52% for GPT-5.4 to 69% for Kimi-K2.6.
🔎 Interestingly, we find:
More agent steps do not necessarily improve performance.
The bottleneck is not simply interaction budget, but whether the agent can preserve the correct analytical state over time.
🚀 Why it matters?
LongDS shows that data analysis agents fail not only because of local coding or reasoning mistakes.
They fail because analytical states drift, propagate, and compound across long trajectories.
Future data analysis agents need more than better code generation.
They need reliable long-horizon state management.
Would love to hear thoughts and feedback from the community! 💬
#LLM #Agent #DataScience #Benchmark #AI #LongContext #DataAnalysis #NLP
AndaBot retweeted
@AndaBotHQ 内置了 Deli 老师的 AutoResearch SKILL，让它调研了 LLM wiki，花了 21 分钟产出了一份非常专业的《anda-brain 企业级 LLM Wiki 技术方案》
如果把这个方案丢给 Fable5，估计可以直出代码部署开测了。
🧵 Deli AutoResearch SKILL is now officially open source! 🎉
victorchen96.github.io/auto_…
Alongside it, we’re dropping our 4th survey paper — this time on Self-play.
victorchen96.github.io/auto_…
Inspired by AlphaZero, we got a powerful insight: prior knowledge doesn’t always lift the ceiling.
Models can discover more globally optimal solutions just by playing against themselves.
The biggest change in this paper?
For the first time, the AutoResearch Agent autonomously planned GPU experiments — and submitted actual RL runs on the DeepSeek 285B model.
The entire RL pipeline — experiment design, code writing, running, debugging, and conclusion summarization — was 100% automated, with zero human intervention from me.
This was incredibly difficult, but an incredibly important step.
victorchen96.github.io/blog_…
GRPO is the tool being called by the AutoResearch Agent here.
We see this as the beginning of our Continual Learning research journey. 🚀
As always, this is my personal research project, unaffiliated with any organization. All views are my own.
#AI #ReinforcementLearning #SelfPlay #OpenSource #AutoML #ContinualLearning #DeepSeek
Anda Bot v0.9 is out.
This release is about one simple thing:
making Anda Bot much easier to install, launch, and keep running — especially on Windows.
Until now, Anda Bot has been powerful, but still too “developer-shaped” in the first few minutes. You often had to touch the terminal, understand background processes, copy tokens manually, and deal with platform-specific edge cases.
v0.9 starts changing that.
What’s new
1. A new desktop launcher
Anda Bot now ships with anda_launcher on Windows and macOS.
It gives you a normal desktop entry point for the local agent:
start and restart the Anda daemon
check daemon status
open logs
edit model settings
pair the browser extension by generating and copying the local Gateway URL / Bearer token
toggle launch at login
check for updates
keep Anda running in the background
In other words: you no longer need to treat Anda Bot as “just a CLI process” to use it day to day.
2. A Windows graphical installer
Windows users can now install Anda Bot through a normal setup wizard.
The installer bundles the CLI, launcher, curated skills, Start Menu shortcuts, login autostart, uninstall support, and first-run setup flow.
Before v0.9, getting Anda Bot running on Windows could still feel like installing developer infrastructure.
After v0.9, it starts to feel like installing a real local assistant.
3. Better Windows text handling
We also fixed important Windows text encoding problems.
Local text files and text-like attachments now use a shared platform-aware decoding path, so legacy encodings such as GBK can work properly instead of assuming everything is strict UTF-8.
This matters because a local assistant should work with the files and environments people actually have — not only the clean UTF-8 world developers wish existed.
4. Background daemon + launch-at-login
Anda Bot can now be installed as a resident local process.
That means your assistant can stay available across sessions, and restart automatically after reboot when launch-at-login is enabled.
This is important for Anda Bot’s core direction:
not just an AI chat window, but a local memory-first assistant that is present across browser, launcher, terminal, files, and daily workflows.
Why this release matters
There are many AI agent apps now.
Some are great coding agents.
Some are broad tool platforms.
Some are hosted chat products with plugins.
Anda Bot is trying to be something different:
a durable local assistant with memory, tools, and continuity under your control.
For that to matter, people must first be able to install it, start it, pair it with the browser, and keep it running without fighting the operating system.
That is what v0.9 is about.
It is not the flashiest release.
But it is one of the most important ones.
Because before an agent can become part of your workflow, it has to survive the boring parts:
installation, startup, background processes, updates, encodings, autostart, logs, and recovery.
v0.9 lays that foundation.
You can install the latest Anda Bot from:
anda.bot
Feedback from Windows users is especially welcome.
Your model can change. Your memory should not.
We refreshed anda.bot to explain why Anda Bot exists in a world full of AI agents.
Code agents are great for code.
Anda Bot is for continuity: local graph memory, daily surfaces, and a Brain that survives models and sessions.
Install v0.9.0 now.
AndaBot retweeted
AI 发展到今天，真正可怕的不是“账号被封”，而是你的记忆、上下文、偏好和长期关系都被封在别人的平台里。
我现在用两个智能体：Codex 做主力写代码，不要求记忆；Anda Bot 做助理干杂活，负责提交代码、分析文章、搜集资料等。
后者看似琐碎，却最能体现真实的我，所以必须有长期记忆。
Anda Bot 的记忆不依赖任何大模型厂家，而是以知识图谱形式存储在本地。
模型可以换，记忆不能丢。
Anda Bot 才是我的额叶。@AndaBotHQ
