For work that lives in documents
Ts-Agents reads your description, derives a team of specialised agents from it, and installs that team in your drive — as markdown files you can open, correct and take with you.
"We publish novels translated into five languages. I want to review the translations without betraying the original, and keep a consistent tone from one volume to the next."
Checks that no fact contradicts the established world
Compares the target version to the source, deviation by deviation
Enforces the register agreed for the series
Models have become very good. What blocks people is elsewhere, and it is always the same thing: between "I know my job" and "AI does it with me", nobody walks you through.
It answers beautifully in general, and poorly on your catalogue, your tone guide, the decision you made three months ago. Every conversation starts from nothing, and you re-explain what you had already settled.
You are asked to write a "system prompt" and to cut your work into roles. What you know is your work — not the specification of a piece of software that ought to understand it.
The agent asks for "the tone notes". They exist nowhere: someone has to write them. It expects "the translations" in a folder nobody created. You are stuck at step zero, with nothing to explain why.
You configure nothing. You explain, you approve, you drop your documents in. Ts-Agents does the rest and shows you every decision before applying it.
In plain language, the way you would explain it to someone joining you. No structure imposed, no vocabulary to learn.
"We publish novels translated into five languages. I want to review the translations without betraying the original, and keep a consistent tone from one volume to the next."
An orchestrator, specialists, and drop folders named in your own words — /manuscripts if you talk about books, /statements if you talk about accounting, /case-law if you talk about legal work. You review, you correct, you apply.
Ts-Agents separates what you must upload from what has to be written together. Folders to fill are created; for documents that do not exist yet — a tone guide, a glossary — the assistant asks you questions and writes them under your control.
A procedure is written into workflow.md: the order of the steps, and above all the moments where the AI must stop and ask for your approval. Full autonomy between two checkpoints, never beyond one.
Ts-Agents knows none of these trades in advance. It derives them from what you describe — which is why the folders are never called /documents.
Review your translations without betraying the original
/manuscripts/translations/<language>/contracts
See the exampleCheck a contract against your own standard clauses, not against the law in general
/contracts/incoming/case-law/models
See the exampleFind the missing document before the client goes looking for it
/statements/<client>/invoices/<client>/filings
See the exampleBid without re-reading 180 pages of tender documents every time
/tender-pack/references/certificates
See the exampleCheck a listing and its surveys before the buyer does
/mandates/surveys/listings
See the exampleAssess against your criteria, not against the impression a CV leaves
/applications/<role>/roles/interviews
See the exampleEngineering practices, architecture firms, training providers, editorial teams, laboratories, grant-managing associations: the criterion is not the sector, it is that your work lives in documents and breaks down into reviews.
Five screens, in the order you meet them.
You describe, it proposes. The list of agents, their speciality and the folders they will need appear before anything is written — nothing is created in your drive until you approve.
"We publish novels translated into five languages. I want to review the translations without betraying the original, and keep a consistent tone from one volume to the next."
Checks that no fact contradicts the established world
Compares the target version to the source, deviation by deviation
Enforces the register agreed for the series
What you have to upload and what still has to be written, kept apart. Each line says what the document is for and which agent will read it. An empty folder is impossible to miss.
The folders are created in your drive. You put your documents in them.
The source texts, one file per volume
The translated text to review, per language
Rights assignments, to check the required mentions
These documents do not exist yet. The assistant asks you questions and writes them.
The register and vocabulary expected for the series
Proper nouns and their fixed translations
You ask a question and watch the organisation work: the file read, the specialist consulted, the document written. You can close the tab and come back — the work carries on.
Does the German translation of volume 2 keep the tone we agreed on?
What the project retained from your exchanges, shown in plain text in memory.md. You can correct a wrong line, or erase everything. A memory you cannot contradict is not one.
Re-read on every request. Correct what is wrong, erase what no longer serves.
## Decisions - Formal address between Elsa and the bookseller, in every language. - Volume 2 is set in winter; do not follow the 2024 manuscript. - Place names are not translated. ## Preferences - Report deviations by chapter, not line by line.
What each exchange cost on your key, per project and per model — input tokens, output tokens, the share served from cache, and an estimate in dollars. Indicative only: your provider sets the price, and only their own count is authoritative.
| Exchange | Model | Input | Cache | Actions |
|---|---|---|---|---|
| Conversation | claude-opus-5 | 12,400 | 9,100 | 6 |
| Conversation | claude-opus-5 | 3,210 | 2,800 | 2 |
| Blueprint | claude-opus-5 | 5,770 | 0 | 0 |
Indicative estimate, based on public rates: only your provider's own count is authoritative.
Faithful reconstructions of the interface, built with the product's real components and colours. These are not screen photographs.
Seven architectural choices, not sales arguments. Each is verifiable in the product in five minutes.
AGENTS.md, one brief per agent, workflow.md, memory.md. Text files you open, read and correct without us. If you cancel, your organisation stays with you — and stays readable.
You plug in your own API key. We do not resell tokens and take no margin on them: your subscription is flat, your consumption is settled between you and your provider.
OpenAI, Anthropic, Gemini, Mistral — your choice, and changeable. No major vendor will offer you its competitors; here, sovereignty is an option like any other.
Checkpoints are written down in workflow.md. Between two of them the AI works alone. At each one, it hands back. You decide where they fall.
The project remembers your decisions from one conversation to the next, and shows you what it retained. A wrong line gets corrected instead of repeated forever.
Folders carry the words of your trade. Talk about books and it will be /manuscripts. Talk about accounts and it will be /statements. Never /documents.
Every exchange is counted, per project and per model, with the number of actions that explains why one request cost ten times another.
There are excellent tools next door. They do not solve the same problem.
| Critère | Ts-Agents | General AI assistant | No-code agent platform | Developer framework |
|---|---|---|---|---|
| Starting point | Your trade, described in plain words | A question | A flow you already know | Code |
| Setup | Proposed, then approved by you | None | Wire it yourself | Write it yourself |
| Where your files live | Your drive, as markdown | The vendor's cloud | The vendor's cloud | Wherever you want |
| Model costs | Your key, no margin | Included, opaque | Resold credits | Your key |
| Switching provider | One setting | Impossible | Sometimes | Recode it |
| Checkpoints | Written in workflow.md | No | To configure | To implement |
| Skill required | Your own trade | None | Technical | Developer |
A comparison of categories, not of named products: their feature sets move too fast for any table to stay accurate for more than a few months.
You pay for the orchestration: the blueprint, the workspace, the workflows, the memory, the consumption tracking. Model calls go through your key and never appear on our invoice.
The plan applies to a project. Open as many as you like, each with its own — a firm can run a free project per client and one Practice project for its document base.
Enough to judge on the evidence, not a crippled demo.
€0forever
When a project becomes a real workbench.
€29per project per month
When several of you work on the same project.
€69per project per month
The questions people ask us before starting.
The blueprint shows you the proposed organisation before writing anything to your drive. Judge it on the evidence.