Obsidian integration is the detail that turns Google’s Gemini desktop app from a chat window into something closer to a workbench. On 5 September 2026, TestingCatalog reported that closed builds of the app carry connections to Obsidian, a macOS Finder action that hands any folder straight to Gemini, and a toggle that splits the app into two modes named Ask and Assign.

None of it has shipped. Everything described below sits in closed testing, and Google has not said when — or whether — any of it reaches the public build. But the shape is unusually legible for a leak, and it points at a product that wants to read the folders on your disk rather than the pages on the web.

That is a meaningful shift for anyone whose working notes live in a local vault rather than a cloud suite, and it is why an Obsidian integration is worth more attention than a typical feature-flag sighting. A large language model that can only see what you paste has a hard ceiling; one that can see a folder does not.

This article covers exactly what the report documented, why the Obsidian integration matters more as a folder story than a note-taking story, how Finder support and the Ask/Assign split fit together, and what teams should weigh before granting any assistant read access to a knowledge base.

What the Obsidian Integration Report Actually Found

gemini desktop obsidian integration finder support b tower of four stacked cubes

The report was published on 5 September 2026 by Bedros Pamboukian and Alexey Shabanov at TestingCatalog, a site that tracks unreleased features in AI clients. It describes a closed build of the Gemini desktop app, not a public release. Every item below is something the authors saw in that build or found in its code.

The mode toggle

The headline change is a toggle between Ask and Assign. Ask presents the standard Gemini interface — the chat window as it exists today. Assign works like Gemini Spark, Google’s agent, but with expanded capabilities. TestingCatalog draws the comparison directly to Chat and Work in OpenAI’s Codex superapp.

The folder scope

Assign lets you choose the folder Gemini operates in. You can set a default, return to a folder used in a previous session, or pick a new one for a specific task. The agent stays inside that scope. This is the mechanism the Obsidian integration would ride on, and it is the part that makes the whole feature set coherent.

The connections

Two integrations stand out in the current testing. The computer use skill can be tagged in the prompt bar to trigger tasks that drive local applications. Separately, Google is testing connections with Obsidian. TestingCatalog is candid that it does not know why: the Obsidian integration could point at a broader knowledge-base concept, or Google may simply use Obsidian internally while building an equivalent of its own.

What was foundWhat it doesEvidence
Ask / Assign toggleSwitches between the chat interface and the agentVisible in the build
Folder scopingDefault folder, previous folder, or a new folder per taskVisible in the build
Remote computer connectionConnect to other machines running Gemini desktopOption present, behaviour unclear
Computer use skillTag it in the prompt bar to drive local appsVisible in the build
Obsidian integrationA connection to the Obsidian note app; purpose unstatedVisible in the build
macOS Finder actionSend any folder directly to Gemini for a taskVisible in the build
Customize tabPick skills, apps, connectors or pluginsVisible, currently limited
Gemini 4 readinessApp being prepared for the next major modelCode traces only
Nano Banana 2.5 FlashAn upgrade to Google’s image modelCode traces only

How firm the evidence is

Seven of the nine items are things a person can see in the client. Two — Gemini 4 and Nano Banana 2.5 Flash — are code traces, which is a weaker class of finding. Separating those two categories is the honest way to read any leak of this kind, and it matters here because the Obsidian integration sits in the stronger group.

The nine findings, split by how firm the evidence is
Seen in the client build — 7 of 9 78%
Code traces only — 2 of 9 22%
Counted from the nine items in the table above; all nine remain in closed testing.

Why an Obsidian Integration Is Really a Folder Integration

gemini desktop obsidian integration finder support c upright panel with two raised blocks

The most useful thing to understand about an Obsidian integration is that Obsidian barely stands between the assistant and the files. Obsidian’s own documentation says notes are stored as “Markdown-formatted plain text files in a vault,” and that a vault “is a folder on your local file system, including any subfolders.”

A vault is a directory

That single sentence explains why the Obsidian integration and the Finder action showed up in the same build. Obsidian does not hold your notes in a database you have to query through an API. It holds them as .md files in a directory, exactly like every other folder on the disk. The documentation goes further and states plainly that you “can use other text editors and file managers to edit and manage notes.”

Which makes the integration cheap

An assistant that can already read a scoped folder is most of the way to reading an Obsidian vault. The remaining work is understanding the conventions — wikilinks, frontmatter, the .obsidian configuration folder — rather than gaining access. That is why an Obsidian integration is a plausible early target rather than an exotic one, and it explains the ambiguity TestingCatalog flagged.

Obsidian propertyWhat the documentation states
Storage formatMarkdown-formatted plain text files
A vaultA folder on your local file system, including subfolders
External editingOther text editors and file managers can edit the notes
Refresh behaviourObsidian picks up changes made outside the app
Multiple vaultsSeparate folders can be opened as separate vaults
ConfigurationA .obsidian folder holds hotkeys, themes and plugins
Sync optionsObsidian Sync, Dropbox, iCloud, OneDrive, Git and others

The multi-vault detail matters

Obsidian users routinely keep separate vaults — one for work, one for personal notes, one per client. Folder scoping in the Gemini desktop app maps onto that habit almost exactly. Pointing an Obsidian integration at one vault while leaving the others untouched is the difference between a useful tool and an uncomfortable one, and the scoping control is what makes it possible.

Finder Support Puts a Folder One Right-Click Away

gemini desktop obsidian integration finder support d strongbox with round dial

The macOS Finder action is the quietest item in the report and possibly the most consequential. TestingCatalog describes it simply: on macOS, Finder integration will let you send any folder directly to Gemini for inspection or a targeted task.

Why the entry point matters

Every local-file assistant so far has made you start in the assistant. You open the app, you attach a folder, you describe the job. Finder support inverts that. You start where the work already is, in the file manager, and push the folder outward. It is a small interaction change with a large effect on how often the feature gets used.

It also normalises the scope

A right-click that says “send this folder” carries its own boundary. The user has picked the directory before the agent has said anything, which is a healthier default than an assistant proposing what it would like to read. Combined with the Obsidian integration, it means a vault can be handed over as a deliberate act rather than a standing permission.

What is still unknown

Nothing in the report says whether the Finder action opens Ask or Assign, whether it inherits the default folder or replaces it, or whether Windows gets the same treatment. Google’s earlier desktop work shipped on both platforms, so a Windows equivalent of the Obsidian integration is reasonable to expect and unwise to assume.

Ask and Assign: The Two Modes Around the Obsidian Integration

gemini desktop obsidian integration finder support e door slab with round knob

The Ask/Assign toggle is the frame the Obsidian integration sits inside, and it is a direct answer to what OpenAI shipped in July. Understanding the split makes the rest of the feature set easier to read.

ConcernGemini desktop (closed testing)ChatGPT / Codex desktop (shipped)
Mode namesAsk, AssignChat, Work, Codex
Conversational modeAsk — the standard Gemini interfaceChat — questions, drafting, light research
Agent modeAssign — Gemini Spark with expanded capabilitiesWork — multi-step deliverables and connected files
Developer modeNone describedCodex — repositories, terminals, local folders
Folder scopingDefault, previous session, or new folder per taskLocal folders via the Codex mode
Note-app connectionObsidian connection in testingNot described in the merge announcement
StatusClosed testing, no public dateMerged into one desktop app from 9 July 2026

Assign is Spark with more rope

TestingCatalog is specific that Assign works like Gemini Spark but with expanded capabilities. Spark already runs continuously in the cloud on Gemini 3.5 Flash using the Antigravity harness, connects to Gmail, Docs and Slides, and asks before high-stakes actions like spending money or sending email. Assign appears to be that agent given a working directory.

The third mode Google has not drawn

OpenAI’s desktop app carries a separate developer mode. Google’s toggle has two positions, not three. Whether the computer use skill and the Obsidian integration end up filling that gap inside Assign, or whether a coding surface arrives later, is one of the more interesting open questions in the report.

Remote machines are the wildcard

The build also carries an option to connect to other computers running Gemini desktop. TestingCatalog does not know how it works and says it will likely resemble the Codex approach. If it means assigning a task on one machine and having it run on another, the Obsidian integration stops being about one laptop’s vault and starts being about a fleet.

How the Obsidian Integration Fits Gemini Spark's Local-File Story

gemini desktop obsidian integration finder support f clipboard with top clip

The Obsidian integration is not a sudden turn. It is the next step on a road Google mapped in public over four months, and reading the sequence makes the direction obvious.

What Google announced in May

Gemini Spark was introduced on 19 May 2026 as “a 24/7 personal AI agent that helps you navigate your digital life” — a shift from answering questions to being “an active partner that does real work on your behalf.” The same announcement said the macOS app would receive Spark that summer for “tasks involving your local files.”

What arrived in June

On 30 June 2026, Google put Spark on macOS in beta for Google AI Ultra subscribers aged 18 and over, starting in the US. The post is explicit about permissions: “Gemini Spark only has access to the files you give it permission to use.” It also announced connected apps and said support for custom Model Context Protocol servers was rolling out.

What surfaced in August

On 18 August 2026, TestingCatalog reported computer use traces in the Gemini desktop settings — a “Task” option in the prompt-bar selector, a mini-view of the active screen, sleep prevention during a task, and an optional automatic backup of the selected folder to Google Drive before work begins. That backup option is the first sign Google expected agents to change files, not just read them.

DateMilestoneLocal-file relevance
19 May 2026Gemini Spark introducedmacOS local-file work promised for the summer
30 Jun 2026Spark macOS beta, US, AI Ultra, 18+Permissioned local-file tasks; custom MCP rolling out
9 Jul 2026Codex merged into the ChatGPT desktop appChat, Work and Codex modes on macOS and Windows
18 Aug 2026Computer use traces in Gemini desktopFolder access, screen mini-view, pre-task Drive backup
5 Sep 2026Ask/Assign, Finder action, Obsidian integrationFolder scoping and a note-vault connection

The pace is the signal

Forty-two days separate the Spark announcement from the macOS beta. Forty-nine more brought the computer use traces. Only eighteen days later, the Obsidian integration and the Finder action appeared. The gaps are getting shorter, which is the clearest available evidence that local-file work is a priority rather than an experiment.

Days between each Gemini desktop milestone
19 May to 30 Jun — Spark to macOS beta 42 days
30 Jun to 18 Aug — beta to computer use 49 days
18 Aug to 5 Sep — computer use to this report 18 days
Bars scaled against the longest gap of 49 days.

The Customize Tab, MCP and the Obsidian Integration Gap

The Customize tab is where the Obsidian integration raises an awkward question: if Google is shipping custom MCP support, why build a bespoke Obsidian connection at all?

What the tab holds today

TestingCatalog describes a Customize tab where users pick skills, apps, connectors or plugins, and calls the current build “quite limited” — mostly Google’s own apps, or extensions already present in the Gemini web interface. A Customize tab was spotted in the desktop client back in August, so this is a surface Google has been building out for a while.

The connector list is Google-heavy

The June update named the connected apps: Google Tasks and Google Keep already worked, and Canva, Dropbox, Instacart, OpenTable and Zillow Rentals were rolling out. That is two Google properties and five third-party services. Obsidian is on neither list, which is what makes a first-party Obsidian integration notable rather than routine.

Connected apps named in the June 2026 Spark update
Third-party services — 5 of 7 71%
Google’s own apps — 2 of 7 29%
Note or knowledge-base apps — 0 of 7 0%
Counted from Google Tasks and Google Keep plus Canva, Dropbox, Instacart, OpenTable and Zillow Rentals.

MCP would make it generic

Custom Model Context Protocol support means any tool that speaks the protocol can be wired in without Google writing code for it. If that lands properly, a dedicated Obsidian integration becomes redundant — which supports TestingCatalog’s second reading, that Google may be using Obsidian internally while building its own knowledge layer. The earlier MCP work in the desktop client points the same way.

Or it is a knowledge-base pilot

The other reading is simpler. Obsidian is the most popular local-first, plain-text note app among the exact users who would try an agentic desktop client first. Shipping an Obsidian integration is a cheap way to find out whether people want an assistant that reads their second brain, before committing to building one.

Gemini Desktop vs Codex: The Obsidian Integration Race

The comparison TestingCatalog makes is not decorative. Google is visibly copying a structure OpenAI already shipped, and the Obsidian integration is one of the few places it is ahead rather than behind.

OpenAI consolidated first

OpenAI merged Codex into the ChatGPT desktop app on 9 July 2026, giving one application three modes — Chat, Work and Codex — across macOS and Windows. Chat handles questions and drafting, Work handles multi-step deliverables and connected files, and Codex handles repositories, terminals and local folders.

Google is answering with two modes

Ask maps onto Chat. Assign maps onto Work. There is no published Google answer to Codex, which is the mode that owns local folders in OpenAI’s design. Instead Google put folder scoping inside Assign, so the Obsidian integration and the Finder action land in the general-purpose agent rather than a developer surface.

That choice has consequences

Putting local files in the mainstream mode makes them available to people who would never open a developer tool. It also means the safeguards have to be right for a much broader audience. An Obsidian integration aimed at note-takers rather than engineers is a wider blast radius, and Google’s pre-task backup option suggests the team knows it.

What an Obsidian Integration Would Mean for Business Teams

For most organisations the specific note app is beside the point. What matters is the pattern: an assistant that reads a scoped local directory, on demand, from the file manager. The Obsidian integration is the visible edge of that.

Knowledge that never reached the cloud

Plenty of genuinely useful institutional knowledge sits in local markdown, meeting notes, exported transcripts and scratch folders — material that was never worth migrating into a document management system. An Obsidian integration is the first credible route to querying that material without a migration project. That is the business case, and it is a real one.

It changes what “connected” means

Today, connecting an assistant to your work usually means connecting it to a SaaS account with an audit trail and an admin console. A folder on a laptop has neither. Any Obsidian integration rollout will force a conversation about which local directories are in scope, and most organisations have never had to answer that question.

The automation angle

The more interesting use is not search but action. A scoped agent that can read a vault, draft from it, and then run a task through the computer use skill starts to look like workflow automation with a much lower setup cost than a traditional integration project. That is also where the governance work belongs.

Where it fits alongside agents

If your organisation is already experimenting with autonomous AI agents, a desktop agent with folder scope is the same class of system with a narrower blast radius and a clearer boundary. The Obsidian integration is worth watching precisely because it is the constrained version of a problem most teams are already facing in a less constrained form.

Obsidian Integration Risks Worth Weighing First

None of this is shipped, which makes it a good moment to think about the risks before a rollout decision is in front of you rather than after.

Scope creep on the default folder

The build lets you set a default folder and reuse it across sessions. A convenience feature that persists is also a permission that persists. If an Obsidian integration defaults to a whole vault, the practical scope is every note ever written, not the ones relevant to the task.

Backups that are not backups

The optional pre-task backup to Google Drive is a sensible safety net and a data-transfer decision at the same time. TestingCatalog raised the practical concern in August: backing up a folder of hundreds of gigabytes could consume substantial storage and time before a task even starts. For a regulated vault, it also moves the data.

Remote machines widen everything

The option to connect to other computers running Gemini desktop is the least-understood item in the report and the one with the largest security surface. Until Google explains it, treat any Obsidian integration on a machine that can be reached remotely as an open question rather than a solved one.

Closed testing is not a roadmap

The most important caveat is the plainest. TestingCatalog says every feature described remains in closed testing with no public release timeline. Features found in builds get cut. Plan for the pattern, not for this specific Obsidian integration.

An Obsidian Integration Readiness Checklist

If you want to be ready to evaluate this properly when it ships, the useful work is not technical. It is deciding in advance what you would allow.

Decide which folders are eligible

Write down, now, which local directories an assistant could read. Most teams find the answer is narrower than expected once it has to be written as a list. That list is what an Obsidian integration should be pointed at, not the whole home directory.

Split your vaults before you need to

Obsidian supports multiple vaults precisely so work and personal notes can live apart. Splitting a mixed vault is much easier before an integration exists than after someone has already connected it. This is the single cheapest preparation available.

Check the tier and the region

Spark’s macOS beta launched for Google AI Ultra subscribers aged 18 and over in the US only. There is no reason to assume the Obsidian integration will be broader at launch. Confirm your licence tier and region before building a plan around it.

Ask the backup question early

If pre-task backups to Drive carry over from the computer use work, decide whether your data is allowed to make that trip. That is a policy answer, not a product one, and it is better settled before anyone enables an Obsidian integration on a working machine.

Watch for the Windows answer

Finder support is macOS-only by definition. If your estate is Windows-first, the interesting question is whether an Explorer equivalent appears. Nothing in the report addresses it. Our AI models and tools hub tracks these releases as they land.

Obsidian Integration FAQ

Has the Obsidian integration been released?

No. TestingCatalog found it in a closed build of the Gemini desktop app on 5 September 2026. Google has not announced it, and the report states that the public release timeline is unclear.

What would the Obsidian integration actually do?

The report does not say. It confirms Google is testing connections with Obsidian and offers two readings: a broader knowledge-base capability, or internal use of Obsidian while Google builds equivalent functionality of its own.

Is Finder support the same feature?

They are separate but related. Finder support is a macOS action that sends any folder to Gemini for inspection or a task. Because an Obsidian vault is just a folder, the two features overlap heavily in practice.

Which platforms get this?

Finder support is macOS-specific. The Ask/Assign toggle, folder scoping and the Obsidian integration were reported without a platform restriction, and Google has previously shipped Gemini desktop clients for both macOS and Windows.

Do I need a paid plan?

Spark’s macOS beta required a Google AI Ultra subscription, US residency and an age of 18 or over. Nothing has been said about the requirements for an Obsidian integration, so assume similar gating until Google states otherwise.

What is Assign mode?

Assign is the agent half of the new toggle. TestingCatalog describes it as working like Gemini Spark with expanded capabilities, including the ability to choose the folder the agent operates in.

Does this replace MCP support?

Unclear, and that tension is the most interesting part of the report. Google announced custom MCP support in June 2026, which would let any compatible tool connect without a bespoke build — making a dedicated Obsidian integration either redundant or a deliberate first-party bet.

References