# Blog

> Writing on team second brains, local-first markdown, real-time collaboration and AI that edits your own files.

Source: https://baalda.com/blog
Site: Baalda, https://baalda.com

- [MCP security risks come from one combination, and a team's notes supply two parts of it](https://baalda.com/blog/is-mcp-safe.md), 2026-10-18. MCP is as safe as the servers you connect to it. The serious risk is one combination: an assistant with your private data, text an attacker wrote, and a way to send data out. Check each server for which of those it supplies and who decides its reach. A team's shared notes bring the first two.
- [Is Obsidian open source? No, and the closed part that matters is the server](https://baalda.com/blog/is-obsidian-open-source.md), 2026-10-16. No. Obsidian is closed source: its release repository says the app's code is not public, though the app has been free for any use since February 2025. Your notes stay markdown files you own, and its API types, Web Clipper and JSON Canvas are MIT. The closed part that matters for a team is Sync, the server holding the shared copy.
- [RAG vs a second brain: why most team notes need links, not a vector index](https://baalda.com/blog/second-brain-vs-rag.md), 2026-10-16. A second brain is the notes a team writes and keeps. RAG is one way to feed notes to a model: chunk them, embed the chunks, retrieve the nearest few per question. For a team vault, an assistant that searches, reads whole notes and follows links is usually enough and easier to check. RAG pays off once the corpus outgrows that.
- [A second brain vs a knowledge base: who the note is for, and where it gets lost in between](https://baalda.com/blog/second-brain-vs-knowledge-base.md), 2026-10-16. A second brain holds notes written for yourself: unfinished, loosely linked, allowed to be wrong. A knowledge base holds pages written for a reader who was not there: finished, owned and kept current. Teams need both, and most lose knowledge at the step between them, so keep both in one place and make promotion a move, not a rewrite.
- [Karpathy's LLM wiki is a second brain for one. Here is what a team adds](https://baalda.com/blog/karpathy-llm-wiki-second-brain.md), 2026-10-16. Andrej Karpathy's second brain is what he calls an LLM wiki: raw sources in one folder, and a wiki of markdown files an AI agent writes and maintains from them, guided by a schema file like CLAUDE.md. For one person, Obsidian, an agent and git are enough. A team also needs live merging, per-folder permissions and undo.
- [Where is claude_desktop_config.json? The path on Windows, macOS and Linux](https://baalda.com/blog/claude-desktop-config-file-location.md), 2026-10-16. On Windows it is %APPDATA%\Claude\claude_desktop_config.json, and on macOS ~/Library/Application Support/Claude/claude_desktop_config.json. Microsoft Store (MSIX) installs on Windows can read a second copy under %LOCALAPPDATA%\Packages\Claude_pzs8sxrjxfjjc\LocalCache\Roaming\Claude instead. On Linux, open it with Settings, Developer, Edit Config.
- [Second brain in Notion: how to build one that still works outside Notion](https://baalda.com/blog/notion-second-brain.md), 2026-10-15. Make one Notion page with sub-pages for Inbox, Projects, Areas, Resources and Archives, keep notes as ordinary linked pages, and use a database only for the project list. Pages export as markdown; full-page databases export as CSV, so a brain built from relations and rollups mostly stays inside Notion.
- [Best documentation tool for teams: who edits it, and where the files live](https://baalda.com/blog/best-documentation-tool-for-teams.md), 2026-10-14. Pick by who edits and where the text lives. Public developer docs: Docusaurus or MkDocs if engineers write in Git, Mintlify or GitBook if others edit too. Internal docs everyone edits: Confluence, Notion, or self-hosted Docmost. Baalda fits when internal docs must stay markdown files a team co-edits live and Git can also read.
- [Build a second brain in Obsidian that more than Obsidian can read](https://baalda.com/blog/obsidian-second-brain.md), 2026-10-14. Make one vault with an inbox and Projects, Areas, Resources and Archive folders. Turn on Daily notes and Templates, give each note a few properties, and link with wikilinks. Keep the answers in the files rather than in plugin queries, so an AI assistant or a teammate reading the folder sees what you see.
- [Knowledge base for startups: the five pages to write before your tenth hire](https://baalda.com/blog/knowledge-base-for-startups.md), 2026-10-14. A knowledge base for a startup is the few written pages that stop the founders being the only place answers live: how the product works, why key decisions were made, how to ship, who the customers are, and a start-here page for new hires. Write those five first, in a tool whose files the company keeps as it grows.
- [Claude and Apple Notes: connecting takes a minute, a bad rewrite is permanent](https://baalda.com/blog/connect-claude-to-apple-notes.md), 2026-10-14. Install Anthropic's Read and Write Apple Notes extension in Claude's desktop app on a Mac and approve the macOS prompt. Claude can then list, read, create and rewrite notes. A rewrite replaces the note's body, and Apple Notes keeps no earlier version of an edited note, so check what Claude changed before trusting it.
- [How to build a second brain that outlives the person who built it](https://baalda.com/blog/how-to-build-a-second-brain.md), 2026-10-13. Build it in four steps: capture only what resonates, organise around your active projects, distill each note in layers, and express the result as something finished. Tiago Forte's CODE method is sound. It assumes one reader, your future self, so a team has to add a fifth constraint: every note legible to someone who was not there.
- [Wiki software is bought for the archive and compared on the editor](https://baalda.com/blog/best-wiki-software.md), 2026-10-13. Pick by who reads it. For a wiki the whole company opens in a browser, Confluence or Notion. Self-hosted, Outline or BookStack. For an archive readable with the app switched off, DokuWiki. For knowing a page is still true, Guru. Baalda fits only when the wiki must be markdown files your team owns.
- [Free Notion alternatives for teams, and the limit each one hits first](https://baalda.com/blog/free-notion-alternative-for-teams.md), 2026-10-13. For three people and no server, AFFiNE's free tier. For a free team wiki you host, Docmost or Outline Community Edition. For Notion-style databases, AppFlowy or AFFiNE. For plain markdown files with live editing, Baalda, free for two on its managed backend and unlimited self-hosted. Notion Free stops at 1,000 blocks once a workspace has 2 or more owners.
- [A collaborative markdown editor usually means a database that accepts markdown](https://baalda.com/blog/collaborative-markdown-editor.md), 2026-10-13. For one shared document in a browser, HedgeDoc. For a self-hosted team wiki, Docmost or Outline. For markdown in an ordinary folder, SilverBullet. Most collaborative markdown editors store text in a database; Baalda is the option with live co-editing and plain .md files on every member's disk.
- [An MCP second brain is only as good as where its notes live](https://baalda.com/blog/mcp-second-brain.md), 2026-10-13. An MCP second brain is one store of notes behind a Model Context Protocol server, so Claude, ChatGPT, Cursor and other clients can search, read and write the same notes. Pick it on three things: plain files that outlive the tool, writes that merge instead of overwrite, and access that follows each person once a team shares it.
- [Obsidian collaboration plugins, compared on whose server carries your edits](https://baalda.com/blog/obsidian-collaboration-plugins.md), 2026-10-13. Relay is the default pick: the most used, free for up to 3 users, with private folders from $6 per user a month. For free and self-hosted, EVC Team Relay has folder roles and Live Share encrypts end to end. Peerdraft is easiest for one-off sessions. All of them add live editing that Obsidian itself does not have yet.
- [Obsidian for teams: what you get, what it costs, and what it can't do yet](https://baalda.com/blog/obsidian-for-teams.md), 2026-10-13. Obsidian for teams is four things: an optional $50 commercial licence, deployment guidance, shared vaults on Obsidian Sync at $4 per user a month, and Publish. A shared vault holds up to 20 people, all with the owner's permissions, and has no live co-editing yet. It suits small teams who edit different notes.
- [Raycast asks before it writes to your notes, not before it reads them](https://baalda.com/blog/raycast-ai-second-brain.md), 2026-10-13. Yes. Raycast AI is an MCP client, so pointing it at a Baalda vault over HTTP gives it your team's markdown notes. Raycast auto-approves read-only tools and asks you to confirm the rest, so the server's own tool annotations decide which calls skip the prompt. Nine of Baalda's twenty-two are read-only.
- [Multi-agent shared memory breaks on the write, not the read](https://baalda.com/blog/multi-agent-shared-memory.md), 2026-10-13. Multi-agent shared memory is one store several agents read and write: a blackboard, a shared block or a database. Reads are easy. The failure is two agents rewriting one entry, where the last write wins and the first is lost. Pick a store whose writes append or target exact text, and that keeps what an overwrite replaced.
- [Transfer ChatGPT memory to Claude, and keep the work the import leaves behind](https://baalda.com/blog/transfer-chatgpt-memory-to-claude.md), 2026-10-13. In Claude, open Settings, Memory, Start import, run the prompt it gives you in ChatGPT, and paste the reply back. That moves a summary of you, not your chats, Project files or anything your team shares. For work that should survive the next switch too, keep it in markdown notes both assistants read over MCP.
- [AI second brain: the three things it means, and which one a team can share](https://baalda.com/blog/ai-second-brain.md), 2026-10-13. An AI second brain is a store of notes that an AI assistant can search, read and add to, so what you learn builds up instead of vanishing with each chat. The phrase covers three things: memory inside one assistant, AI built into one notes app, and notes you own that any assistant connects to over MCP.
- [Connecting Claude to Notion hands it every page you can open](https://baalda.com/blog/can-claude-connect-to-notion.md), 2026-10-12. Yes. Notion hosts an MCP server at mcp.notion.com/mcp, and one OAuth grant lets Claude search, read and write your pages. Notion's help page states that MCP tools act with your full Notion permissions and can access everything you can access, so what Claude reaches is your whole workspace scope, not a page you picked.
- [Best Notion alternatives in 2026](https://baalda.com/blog/best-notion-alternatives.md), 2026-10-12. The best Notion alternative depends on why you are leaving. For owned markdown files with live co-editing and AI over MCP, Baalda. For an open-source swap that keeps databases, AppFlowy or AFFiNE. For encrypted local-first spaces, Anytype. For a team wiki, Outline or Confluence. For docs that behave like apps, Superhuman Docs, formerly Coda.
- [Best Obsidian alternatives in 2026](https://baalda.com/blog/best-obsidian-alternatives.md), 2026-10-12. If your notes must stay markdown files you own, start with Baalda for a team (live co-editing, folder permissions, AI over MCP), Logseq OG for a free open-source outliner, or Reflect Open on a Mac. Joplin and Anytype give up the files for free sync or live sharing, Notion for a cloud workspace. Working alone and happy? Stay on Obsidian.
- [Note taking apps for teams do one of two jobs, and the lists mix them up](https://baalda.com/blog/best-note-taking-app-for-teams.md), 2026-10-12. Two different products share the phrase. If nobody is capturing meetings, buy an AI notetaker: Otter, Fireflies, or Teams' own recap. If the team cannot find what it decided, buy a shared knowledge base: Notion, OneNote, ClickUp Docs, or Baalda for markdown files you own with live co-editing. Most teams need the second.
- [Codex CLI will remember your preferences and forget your team's decisions](https://baalda.com/blog/codex-cli-second-brain.md), 2026-10-12. Codex CLI stores its memories under `~/.codex/memories/`, off by default and generated by Codex from its own past threads. They live on one machine, under one account, and hold your preferences rather than your team's decisions. Connecting Codex to a Baalda vault over MCP puts the durable half in markdown files the whole team can open.
- [Owning the server is the easy half of self-hosted shared notes](https://baalda.com/blog/self-hosted-shared-notes.md), 2026-10-12. Self-hosting a notes app gives you the server, not multi-user. "Shared" means four different things: replicated to your own devices, published as a link, granted to another account, or co-edited live under a permission. Most self-hostable apps do the first three, and the ones that co-edit properly usually stop storing your notes as readable files.
- [n8n can write into the same markdown files your team is editing](https://baalda.com/blog/n8n-second-brain.md), 2026-10-12. Yes. n8n's MCP Client Tool node is an MCP client, and Baalda serves the Model Context Protocol at POST /api/mcp over Streamable HTTP. Set Server Transport to HTTP Streamable, authenticate with a Bearer token minted in Baalda, and the AI Agent node can read, create and append to your team's markdown files.
- [LibreChat over MCP: self-host the chat, then self-host what it writes](https://baalda.com/blog/librechat-second-brain.md), 2026-10-12. Yes, LibreChat is an MCP client. Add Baalda under `mcpServers` in `librechat.yaml` as a `streamable-http` server pointing at `POST /api/mcp`, with each user supplying their own vault token through `customUserVars`. The assistant then reads and writes the same markdown files your team edits, within that person's own permissions.
- [Does the context window reset? It is rebuilt from nothing every time](https://baalda.com/blog/does-context-window-reset.md), 2026-10-11. No. The context window keeps nothing between requests, so there is nothing to reset. Your client replays the whole transcript on every turn, and the model holds none of it afterwards. What people call a reset is a new session, compaction replacing old turns with a summary, a manual clear, or a usage allowance. Only the first is reversible.
- [Three ways to run an MCP server for Obsidian, and what each one needs switched on](https://baalda.com/blog/mcp-server-for-obsidian.md), 2026-10-11. Three routes exist. The Local REST API plugin now ships a built-in MCP server at /mcp/ on 127.0.0.1, authenticated with a bearer token. Wrapper servers like mcp-obsidian call that same plugin. Filesystem servers read the .md files directly. The first two need Obsidian open; all three hold one key over one vault.
- [Gemini CLI loads GEMINI.md into every request. Your team's notes belong behind a tool call](https://baalda.com/blog/gemini-cli-second-brain.md), 2026-10-11. Gemini CLI keeps what it knows in GEMINI.md, which it concatenates into every prompt and stores on one machine. Connecting it to a Baalda vault over MCP moves the team's notes behind search_notes and read_note, so they are fetched when needed rather than carried on every request, and written back through anchored edits.
- [The best memory tool for Claude Code depends on what actually forgot](https://baalda.com/blog/best-memory-tool-for-claude-code.md), 2026-10-11. Claude Code already has two memory systems: CLAUDE.md, which you write, and auto memory, which it writes to ~/.claude/projects/<project>/memory/. Most of the tools ranked for this question replace one of those for a single developer. The case none of them covers is memory a whole team has to read, and that only some of them should.
- [Logseq as a team knowledge base means choosing which Logseq you mean](https://baalda.com/blog/logseq-team-knowledge-base.md), 2026-10-11. Logseq split into two apps in April 2026. Logseq OG keeps every page as a markdown file but has no multi-user mode and is now maintained rather than developed. Logseq 2.0 adds real-time collaboration and stores the graph in db.sqlite instead of files. A team is choosing between owning the files and having more than one person in them.
- [Onboarding documentation is tested once per hire, by the person least able to fix it](https://baalda.com/blog/onboarding-documentation.md), 2026-10-11. Write it as a checklist of the first week's real tasks, then do the part the templates skip: give the newest person edit access to it on day one. Onboarding documentation is tested properly once per hire, by the person who has just hit every gap in it and usually cannot change a line.
- [Can Obsidian be a team knowledge base? Only if everyone reads everything](https://baalda.com/blog/obsidian-team-knowledge-base.md), 2026-10-10. Yes, for a team where everyone may read everything. Obsidian Sync shares a whole vault with up to 20 collaborators who all get the owner's permissions, and Publish protects a site with one password and explicitly no per-note protection. Neither can give one person access to one folder.
- [Give an AI agent memory that a person can open and correct](https://baalda.com/blog/give-ai-agent-memory.md), 2026-10-10. Store it outside the model and feed it back each run. The usual answer is a memory framework that extracts facts into its own database. For a team, make the memory a markdown note instead: the agent amends it with anchored edits over MCP, and a person can open the same file and fix a wrong line.
- [Why Obsidian sync conflicts happen, and why merging afterwards is too late](https://baalda.com/blog/obsidian-sync-conflicts.md), 2026-10-10. A conflict happens when the same file changes on two devices before they sync. Obsidian merges markdown with Google's diff-match-patch, which applies patches best-effort, so its own docs warn the result may contain duplicate text. Merging two finished files is a guess. Merging edits as they are typed is not.
- [Self-hosting Obsidian notes is three different jobs people keep confusing](https://baalda.com/blog/self-host-obsidian-notes.md), 2026-10-10. Obsidian has no server to self-host. Its own help calls a vault "a folder on your local file system", so self-hosting means running one of three separate things beside it: a sync backend such as CouchDB for Self-hosted LiveSync, a desktop in a container, or a published website. None of them adds users or permissions.
- [An MCP server for team notes has to know who is asking](https://baalda.com/blog/mcp-server-team-notes.md), 2026-10-10. One that resolves permissions per person, not per installation. Most note MCP servers use a single API key and a folder allowlist, so every caller acts as the vault. For a team, the server has to identify the caller and re-check what that user may read on each tool call, against the same ACL the app uses.
- [Pick a second brain for Claude on the write path, not the read](https://baalda.com/blog/best-second-brain-for-claude.md), 2026-10-10. For one person, the best second brain for Claude is a markdown folder with Claude Code pointed at it: its Edit tool does exact, unique string replacement rather than rewriting the file. That answer runs out at the second person, where you need a store that serialises writes, checks a revision, and refuses a stale one.
- [Document knowledge transfer where the work already is, not in a handover pack](https://baalda.com/blog/document-knowledge-transfer.md), 2026-10-10. Document it twice over. Write decisions, runbooks and open threads into the team's own files as the work happens, so the record exists before anyone resigns. Then move the access rather than the content: notes that already live in a shared vault need no migration on the last day, because only the membership ends.
- [How to write an AGENTS.md file, and what to leave out of it](https://baalda.com/blog/how-to-write-agents-md.md), 2026-10-09. AGENTS.md is plain markdown with no required fields, read by 23 tools listed on agents.md. Write only what has to be true for every session in that directory: build and test commands, project layout, the conventions the agent keeps breaking. Codex concatenates every file from the repo root down and stops at 32 KiB, so length is a shared budget.
- [A Claude Code second brain for a team: what breaks when the second person arrives](https://baalda.com/blog/claude-code-second-brain-for-a-team.md), 2026-10-09. Point Claude Code at a folder and you have a second brain for one person. For a team you need three more things: the notes reachable from every machine, concurrent writes that do not duplicate text, and per-folder permissions the agent inherits. Baalda supplies those over one MCP endpoint.
- [Confluence alternatives that use markdown, and where each one actually puts the file](https://baalda.com/blog/confluence-alternative-markdown.md), 2026-10-09. Confluence stores pages as XHTML-based XML and exports PDF, CSV, HTML or XML, never markdown. So judge an alternative by where it puts the file, not by whether it has a markdown editor. Outline, BookStack and Docmost keep yours in a database. Wiki.js and Baalda write real .md files to disk.
- [ChatGPT makes a good second brain and a bad place to keep one](https://baalda.com/blog/chatgpt-second-brain.md), 2026-10-09. Two ways, and they do different jobs. Custom instructions, Projects and uploaded files make ChatGPT behave like a second brain, scoped to your account. Connecting it to a custom MCP server makes it the interface to one, reading and writing markdown files you keep. Use the first for private thinking, the second for anything a team relies on.
- [Claude Code does have memory now. It just never leaves your machine](https://baalda.com/blog/does-claude-code-have-memory.md), 2026-10-09. Yes, two kinds. CLAUDE.md files you write, and auto memory Claude writes itself into `~/.claude/projects/<project>/memory/`, both loaded at the start of every session. Anthropic's documentation states that auto memory is machine-local and not shared across machines, so five teammates on one repository keep five separate memories of it.
- [Internal documentation: write the decisions, index the files you already have](https://baalda.com/blog/internal-documentation.md), 2026-10-09. Write the things only a person knows: the decision, the reason, the convention, the exception. Do not retype the spreadsheet, the contract or the exported report. Put those where one search reaches their words as well as your notes, under the same permissions, and keep the index tied to the bytes it describes.
- [Moving a team from Notion to Obsidian without losing the multiplayer half](https://baalda.com/blog/move-from-notion-to-obsidian.md), 2026-10-09. Export your Notion workspace as HTML, not Markdown: Obsidian's own import docs say the Markdown export omits important data. Then run the Importer plugin, choose Notion (.zip), and point it at the file. Connecting your Notion account instead preserves databases, converted to Bases. Neither route brings your teammates across.
- [Obsidian per-file permissions need a server that can say no](https://baalda.com/blog/obsidian-per-file-permissions.md), 2026-10-08. Obsidian has no per-file permissions. Its own help page says every collaborator on a shared vault "receive the same permissions as the vault owner". Because a vault is a folder on disk, a permission needs a server that can refuse to send a file. VaultGuard Sync, Vault Rooms, Relay and Baalda each add one.
- [Cursor can now reach the agent on your laptop. Your teammates still cannot](https://baalda.com/blog/remote-control-local-agent-notes.md), 2026-10-08. Reachability moved, durability did not. Cursor's Remote Control, shipped 6 October 2026, lets the iOS app message agents running on your own computer, which must stay awake. The written record of a run still lives on that one machine. Giving the agent a shared vault over MCP is what changes whose disk it lands on.
- [How to share an Obsidian vault with others, free and paid](https://baalda.com/blog/share-obsidian-vault-with-someone.md), 2026-10-08. Four ways. Obsidian Sync shares the whole vault, but every collaborator needs their own subscription and gets the owner's permissions. Publish is read-only. A shared folder or Git copies files and loses simultaneous edits. The Relay plugin adds live editing. To share one folder with one person at a chosen level, you need per-file permissions.
- [The team's vault is the one MCP server that should not run on a laptop](https://baalda.com/blog/local-mcp-server-your-team-vault.md), 2026-10-07. On the server your team runs, not as a process on each laptop. MCP's Local Server Security guide, merged 5 October 2026, says a local stdio server inherits your whole account and that folder allowlists enforce nothing. A vault reached over authenticated HTTP installs nothing on anyone's machine.
- [Agent memory fails by throwing things away, not by searching badly](https://baalda.com/blog/agent-memory-retention-not-retrieval.md), 2026-10-07. Because it was deleted, not mislaid. A benchmark published 30 September 2026 found every failure under bounded recency memory was an eviction, none a ranking error. Better search cannot surface a fact the memory already dropped, so what a team keeps has to live somewhere nothing ages out of.
- [Google open-sourced the embedder. Your notes app kept the index](https://baalda.com/blog/open-embedder-your-search-index.md), 2026-10-07. An embedding index is derived data, so the team that owns the source files can rebuild it with any model. Google released EmbeddingGemma 2 under Apache-2.0 on 6 October 2026. It only helps your team's knowledge if the notes are files you hold and the indexer is a component you can replace.
- [Mistral Large 4 scored 59.9% on business workflows someone else had already written down](https://baalda.com/blog/who-writes-the-workflow-down.md), 2026-10-07. Only where the work is already written down. Mistral Large 4, in public preview since 6 October 2026, scores 59.9% on AutomationBench across 657 business workflows that the benchmark specifies for it. Inside a company nobody specified them, so the fix is written context the agent can read and add to.
- [Research says a model's context belongs in a file. Keep one your team can open](https://baalda.com/blog/context-file-teams-can-open.md), 2026-10-06. In a file the team keeps. A UW and Meta paper published 29 September 2026 found a model manages its context better as a file it rewrites than as an append-only transcript. That file is scratch and dies with the run, so the durable version has to be a markdown note on your own disk.
- [When your team's knowledge is a person, an AI has nothing to cite](https://baalda.com/blog/nothing-to-cite-team-knowledge.md), 2026-10-06. Nothing, unless the knowledge is already a file. Stack Overflow's 2026 Developer Survey, published 6 October 2026, found 93% need source attribution before trusting AI output, while the top source for work answers is a coworker at 72%. Citable team knowledge means notes at real paths an agent can read and a person can edit.
- [Open weights, rented notes: what Beam's Apache-2.0 licence does not cover](https://baalda.com/blog/open-weights-self-hosted-notes.md), 2026-10-06. An Apache-2.0 model licence covers the weights and nothing else. If you self-host Beam but your team's knowledge sits in a vendor's database, you moved the compute inside your perimeter and left the corpus outside it. Owning the stack means owning both: the model you run and the files it reads.
- [Nobody keeps a list of what is subscribed to your team's notes](https://baalda.com/blog/what-is-subscribed-to-your-notes.md), 2026-10-06. The one serving them. OpenAI shipped ChatGPT support for MCP Events on 29 September 2026, so a server can push data on a standing subscription, and the draft defines no way to list live subscriptions. Whatever holds your notes has to keep that list itself: every connection, what it reached, and a revoke button.
- [Your notes outlive the model that reads them](https://baalda.com/blog/notes-outlive-the-model.md), 2026-10-06. Keep them in files you own. Anthropic deprecated Claude Sonnet 4.5 on 30 September 2026 for retirement on 30 November, and OpenAI deprecated three models a day later. Models churn on a 60-day to 6-month clock. Plain markdown on your own disk, reached by any AI over MCP, has no retirement date.
- [Gemini 4 Argon: benchmarks, price, release date, and where a million tokens of output lands](https://baalda.com/blog/gemini-4-argon-where-the-output-lands.md), 2026-10-05. Gemini 4 Argon can emit up to 1 million tokens in one response, roughly fifteen times its predecessor's 64,000. That moves the bottleneck from generating work to checking it. Output that lands in a chat transcript cannot be reviewed by a team, so it has to land in shared files several people can edit at once.
- [Mods made Claude Code's instruction loader replaceable. Your team's conventions never fitted one repo](https://baalda.com/blog/claude-code-mods-team-instructions.md), 2026-10-05. Where should a coding agent's instructions live for a team? Not in one repository. Claude Code's docs list exactly one team-shared location for project instructions, a `CLAUDE.md` shared through source control, and the auto memory Claude writes itself is machine-local. Conventions spanning several repos belong in a shared vault, read over MCP under per-folder permissions.
- [Before you give an agent write access, take a checkpoint](https://baalda.com/blog/agent-write-access-checkpoint.md), 2026-10-05. Take a labelled checkpoint of the knowledge it will touch, before it has write access. BCG found 42% of companies expect agents to decide without approval by 2030 and only 5% have the controls. A whole-vault checkpoint you can roll back is the one control a small team can install this week.
- [Windsurf is Devin Desktop now. The project memory problem came with it](https://baalda.com/blog/windsurf-second-brain.md), 2026-10-04. Yes. Devin Desktop, the editor Windsurf became in June 2026, is an MCP client. Its config still sits at `~/.codeium/windsurf/mcp_config.json`, and a remote server takes a `serverUrl` with a `headers` map. Point it at a Baalda vault's `/api/mcp` endpoint and the agent reads and writes your team's markdown.
- [Shared agent memory shipped at DevDay. The open question is whose disk it is on](https://baalda.com/blog/shared-agent-memory-whose-disk.md), 2026-10-04. In files your team owns. OpenAI's dots form their own memories, and OpenAI's FAQ says you cannot view, delete or directly modify an individual one; resetting deletes the whole agent. A vault of plain markdown inverts that: the agent writes notes over MCP, so correcting what it knows means editing a line.
- [Zed's context servers can reach the notes your whole team is editing](https://baalda.com/blog/zed-second-brain.md), 2026-10-02. Yes, through a context server. Zed reads MCP servers from the `context_servers` key in `settings.json`, and a remote one takes a `url` and a `headers` map. Point it at a Baalda vault's `/api/mcp` endpoint with your own token, and the agent panel reads and writes the same markdown your team edits.
- [Goose forgets the middle of a long session. Its memory belongs in a vault](https://baalda.com/blog/goose-second-brain.md), 2026-10-01. Yes, as a remote extension. Goose supports `streamable_http` extensions, Baalda serves a vault at `/api/mcp`, and `goose configure` wires them together with an Authorization header. The agent then reads and writes the team's markdown under your own permissions, so what it learns survives compaction and the session ending.
- [A memory bank Cline can reach from any repo](https://baalda.com/blog/cline-second-brain.md), 2026-09-30. Yes, over MCP. Cline is an MCP client, Baalda serves a vault at `/api/mcp`, and one entry in Cline's MCP settings connects them. Set the transport to streamableHttp explicitly: Baalda answers POST only, so the SSE default will not connect. The agent then reads and writes the team's markdown under your own permissions.
- [The Claude Desktop second brain, without touching a config file](https://baalda.com/blog/claude-desktop-second-brain.md), 2026-09-29. Yes, over MCP, and the route depends on where the notes live. A vault on a server connects as a custom connector: paste the URL into Claude Desktop's settings and sign in. The local config file is only for MCP servers that run on your own machine as a child process over stdio.
- [Claude Code's second brain is a folder, not an integration](https://baalda.com/blog/claude-code-second-brain.md), 2026-09-28. Yes, and for a local vault there is nothing to install. Claude Code edits files on your disk directly, so pointing it at your Baalda folder is the whole setup. MCP is for the other case: a vault on a server, an agent running elsewhere, or a token scoped tighter than your own filesystem access.
- [Cursor as a second brain: the notes that explain your repo](https://baalda.com/blog/cursor-second-brain.md), 2026-09-26. Yes. Cursor is an MCP client and Baalda exposes your vault as an MCP server, so the two connect. Add Baalda to .cursor/mcp.json with a vault token, and the agent that already knows your codebase can also read and write the decisions, specs and postmortems that explain why the code is the way it is.
- [Hermes Agent needs a second brain. Here is how to connect one](https://baalda.com/blog/hermes-second-brain.md), 2026-09-14. Yes. Hermes Agent ships with a built-in MCP client and Baalda exposes your vault as an MCP server, so the two connect. Add Baalda to ~/.hermes/config.yaml with a vault token, and Hermes reads and writes the same markdown files your team edits, instead of keeping memory in a store only it can read.
- [OpenClaw with a second brain: an always-on agent your team can edit](https://baalda.com/blog/openclaw-second-brain.md), 2026-09-09. Yes. OpenClaw is an MCP client and Baalda exposes your notes as an MCP server, so you can connect the two. Add Baalda to OpenClaw's config with a vault token, and your always-on agent reads and writes the same markdown files your team edits, instead of keeping memory in a private silo.
- [Does Obsidian have live collaboration? Not yet, and what teams use until it does](https://baalda.com/blog/obsidian-multiplayer-realtime-collaboration.md), 2026-08-27. No. Obsidian's help says it "does not yet support collaborative live editing on the same file", and multiplayer is on its roadmap with no date. Shared vaults on Obsidian Sync merge edits after they sync. For live cursors today, use a plugin such as Relay, or a markdown app built for teams like Baalda.
- [What is a team second brain (and why nobody built one until now)](https://baalda.com/blog/team-second-brain.md), 2026-08-27. A team second brain is a shared knowledge system a whole team reads and writes together, instead of each person keeping private notes. It combines linked notes and capture with real-time collaboration and shared ownership, so the team's collective knowledge lives in one place everyone can use.
- [The open source second brain: own your notes, forever](https://baalda.com/blog/open-source-second-brain.md), 2026-08-26. An open source second brain keeps your notes in a format you can read, own, and self-host, with no vendor lock-in. Baalda is open source under Apache-2.0: plain markdown files on your disk, self-hostable, and free locally, with real-time team collaboration and AI over MCP on top.
- [Connect Claude to your notes: a second brain over MCP](https://baalda.com/blog/mcp-notes-ai.md), 2026-08-25. MCP (Model Context Protocol) lets an AI assistant read and write your actual note files instead of a separate chat window. With Baalda, notes are exposed over MCP so an assistant like Claude edits the same markdown you do, and its changes merge with yours in real time through the same sync.
- [The best second brain apps in 2026 (for teams and solo)](https://baalda.com/blog/best-second-brain-apps-2026.md), 2026-08-24. For one person, Obsidian and Reflect Open lead on owned local markdown, Logseq if you want it free and open, Notion for an all-in-one cloud workspace. For a team that needs several people editing the same owned markdown at once, with permissions and AI over MCP, Baalda is the strongest fit.