Skip to content
TenhariRequest access

The fine print: what it does, and where it stops

Everything Tenhari does today and everything it does not do yet, in full. We would rather tell you where the edges are than have you find them during a client call.

The landing page says the short version. This page is the long one, and the two never disagree: when the product changes, both change in the same step.

Last updated 21 November 2026.

Where it runs

Tenhari runs in two places, and they do different jobs. The hosted instance is where review happens: boards, pins, notes, screenshots and voice recordings are saved in a server database we host ourselves, and a share link opens on any device.

The same product also runs on the machine that holds your repository. That is where the agent works: the run, the diff, the proof, the commit, the merge, the push, the undo and the before and after re-shoot all happen there. Run locally with no environment configured, it keeps its boards in the browser instead of on a server.

It runs Claude Code and Aider for you, on the machine that holds the repository. Any other agent takes the exported tasks.md file or the copied prompt, or connects over MCP to read the pins and hand each one back for your review.

Everything that works today

Voice notes

Say the note on the pin instead of typing it. Chrome, Edge and Safari transcribe it through the browser's own speech service, and the recording stays with the note.

Proof on each fix

After a run in the app, each pin it worked on comes back with its spot before and after at Desktop, Tablet and Phone on a live page, the files and lines the run changed for it, kept only where the run's own diff agrees, and up to 40 other pins re-shot and compared with their stored screenshots.

Your review of each fix

Approve and the pin resolves; reopen and it returns to the queue with your reason in its thread. The run's exact changes stay in the owner-only What changed view, since anyone on the share link can read that thread. A clean run is not a correct change, so a person still decides.

Client sign-off

On a comment link, a reviewer sees each change the team approved, before beside after at the screen sizes it was checked at, and approves it or asks for changes with a note. Asking reopens the pin, and the team sees who signed off and when.

Approve to merge

A run can make its change on a branch of its own. Once your team and the client have approved every pin the branch carries, the machine that holds the repository merges it and pushes: by your hand, or by itself where you switched that on, once the last approval has stood for 15 minutes.

Questions back to the client

When a note is unclear, the team asks the reviewer a plain question on the pin, with answers to tap. Your coding agent can draft one; a person on your team asks it. The answer lands under the question, and your agent can read it.

Run, diff, push and undo in one place

On the machine that holds the repository, the agent run against the board's folder, the diff, the push and the undo all sit beside the board.

Your own agent, over MCP

With an API token you create and can revoke, a coding agent on your machine reads a board's pins and writes back: a team-only note, the pin handed back for your review, or a question for the client drafted for your team to ask.

Pins on the live page

A note is anchored to its spot on the rendered site and finds its element again when the content above it changes, or sits on a screenshot when a page will not render.

Drawn markup

Circles, arrows and text drawn on a pin's screenshot travel with the task, flagged to the agent as part of the instruction.

Whole-site boards

One board covers a site's pages, pulled from its sitemap where the site exposes one.

Share links for reviewers

A link opens on any device. The reviewer types a display name and needs no account. Read or comment scope, a 30 day expiry, and a revoke.

The tasks.md export

Each pin becomes one task in a single tasks.md file, with its screenshots, as a zip or as a prompt to copy. A test locks the format so it cannot drift.

One queue across boards

Open pins from every board queue up in one list, so a week of notes is not eight tabs.

Activity log and webhooks

On the hosted instance the server records every change to a board in its activity log, and signed webhooks send each event to an address you choose, Slack included.

Accounts for your team

Everyone on your team signs in as themselves, invited by an admin. An account holds an email address, a name, a role, a hash of the password, and the browsers signed in with when each was last used. Admins see the team and can remove anyone at once.

A workspace for each team

Each team works in its own workspace, kept apart from every other team's on the same server. An admin can download everything the workspace holds as one file, or delete the workspace for good.

Self-hosted storage

The hosted instance keeps its data in a database on our own host.

What is not there yet

You start each run

Notes become tasks by themselves. The agent's run does not start by itself yet: today you start each run with a click, on your own machine.

You look before your client does

After a run, each fix comes to you first, with its before and after. Your client is asked to approve only once you have approved it yourself.

Messages to your client

The message that tells a client their change is ready is an email, and the hosted site does not have mail set up yet. Until it does, you tell them yourself and they open the same link.

Agent runs on the hosted site

The hosted site does not run an agent or connect to a repository. It gives you the export, a prompt to copy, and an endpoint for a coding agent on your own machine: with an API token you create and can revoke, the agent reads a board's pins and hands each one back for your review. A run that edits files happens on the machine that holds the repository, in the product or in your own agent, and the model providers reached over HTTP answer in chat without editing files. The product running there can also open the hosted board itself with a token, so the run, its diff and your review happen on that machine and the status lands on the board your client sees.

Sign-up, plans and prices

You cannot sign up on your own yet: we create each team's workspace and send its first invite. A workspace is on a plan that can cap its boards, its people and its stored files, but the tiers and their prices are not set, so none is published and payment is not switched on.

Merging by itself

A branch merges by itself only on the machine where you switched that on, and never in a repository that has merge or push hooks. Everywhere else you merge by hand from the queue.

Pin drift

A pin finds its element again through a small script in the page, which the review proxy adds for you. A site framed directly, without the embed script, cannot run it: there a pin keeps its position on the page, and after the content above it changes length a marker can end up beside the thing it pointed at.

Pages that will not render

Heavy client-rendered sites may not render inside the review frame. The fallback is a screenshot board: pin a full-page screenshot instead.

Transcription

Chrome, Edge and Safari send a voice note's audio to the browser vendor's speech service. Other browsers keep the recording without a transcript unless the operator has configured a transcription service.

The questions, in full

What is Tenhari?

Tenhari is a self-hosted website feedback and annotation tool. A reviewer pins a note on the live page by talking, typing or drawing, and every pin becomes a structured task in one tasks.md file that a coding agent can read. It runs Claude Code and Aider on the machine that holds the repository, and each fix comes back with its proof: before and after at three screen sizes, the files and lines it changed, and a check of the other pins, for you to approve and your client to sign off.

How is it different from a comment-only feedback tool?

A comment-only tool is finished when the note is filed. Here the note is where the work starts and the client's sign-off is where it ends. The client says what they mean on the spot they mean; the note becomes a task an agent reads; the agent runs against your repository on your own machine; the change comes back re-shot at three screen sizes, with the files and lines it changed and a check of the other pins; the client approves it, or sends it back with a note, from their own link; and the approved change can merge from your machine.

Which coding agents does it work with?

Claude Code and Aider edit files from inside the product, on the machine that holds the repository. Any other agent can take the exported tasks.md file with its screenshots or the prompt you copy from the board, or connect over MCP with an API token, read a board's pins and hand each one back for your review. The model providers reached over HTTP answer in chat and do not edit files.

The tasks.md format

How automatic is it today?

Most of the way. A note becomes a task by itself, your coding agent makes the change without you retyping anything, and every fix comes back with its before and after. Three things still take a person today: you start each agent run with a click, you look at each fix before your client is asked, and you tell your client it is ready, because the hosted site does not send email yet.

Does a client's approval merge code by itself?

Only where you switched that on, on the machine that holds the repository. A run can make its change on a branch of its own. Once your team and the client have approved every pin the branch carries, you can merge and push it by hand from the queue. If you opted in on that machine, it merges by itself once the last approval has stood for 15 minutes, and until then the board shows which branch is about to merge, with a way to merge it now or stop it. A repository with merge or push hooks never merges by itself.

Can I see it working before I ask for access?

Yes. The sample board is a board on an invented practice's site, shown the way a client's reviewer sees it: pins on the page, a spoken note, a question with its answer, and a change before and after at three screen sizes. It needs no account, and nothing you click on it is saved.

Open the sample board

Is this live, or a waiting list?

Live, and we use it. The hosted instance stores boards in a server database and its share links open on any device. What the hosted site does not do is run an agent: a run that edits files happens on the machine that holds the repository, in the product or in your own agent. Access goes through a conversation because that split matters and a form cannot explain it.

Do you need access to my repository?

No. The hosted site never connects to a repository. It offers the export, a prompt to copy, and an endpoint your own agent can read a board's pins from with an API token you create. A run that edits files happens on the machine that holds the repository, so your code goes to the model provider behind the agent you chose and not to us. What that agent sends back is a short note on each pin it worked on.

Does it work on a site I do not host?

Often. Marketing sites, docs, and landing pages usually render inside the review frame. Heavy client-rendered sites may not. When a page will not render, make a screenshot board and pin a full-page screenshot instead.

Where does my client's data go?

On the hosted instance, the names reviewers type, their notes, screenshots and voice recordings are saved in a self-hosted database on our own host, and they stay there until the board is deleted, and for 30 days after that, while the operator can still restore it. Everyone holding a board's share link can see what is on that board. A voice note recorded in Chrome, Edge or Safari is transcribed by the browser's speech service, which sends the audio to the browser vendor; other browsers keep the recording without a transcript unless the operator has configured a transcription service.

Privacy and data, in full

Does my client need an account?

No. They open the share link on any device, type a display name and start. A link has read or comment scope, lasts 30 days unless the owner picks another length, and can be revoked. Accounts are for your own team: each person signs in with an email address and a password after an admin invites them, and the team works in its own workspace, kept apart from every other team's.

What does it cost?

Nothing extra on a Fractional CTO retainer, where Tenhari is included. No pricing table is published because standalone pricing does not exist yet, and we would rather publish nothing than a number we already intend to change.