# Buff Code > A development environment for engineers who write their own code: the company’s models, references and trackers, shared by the team, billed to the organisation. Buff Code # An agent in your project folder — *under your control* An app and a terminal program for engineers: the agent reads code, edits files and runs checks in your project folder. You decide what it may do, see every edit and roll it back with one button. [Download Buff Code](https://buff.systems/en/downloads#buff-code) [See it in the demo](https://buff.systems/demo-app/auth/demo?to=-%2Fdev-sessions) [How to work with it](https://buff.systems/en/docs/buff-code/work) There is nothing extra to pay — usage is charged to the organisation’s balance **Buff Code** app ← scroll the frame → **We asked for a free delivery function and a test for it:** the agent wrote them and ran the checks. Under the answer — what changed and how to undo it. Objection 01 ## “The agent will make edits I won’t have time to check” Especially when it “just quickly fixes this bit too”. ### Every edit is visible and can be rolled back A turn is one request of yours and everything the agent did in response. You decide, for each session, how much the agent may do without asking. Everything it did in a turn is gathered under the answer: which files and lines changed. Don’t like it — roll back the whole turn; Buff Code shows in advance what will be restored. #### You decide how much it may do without asking Three modes: “Ask first” — before every edit and every program run; “Edit on its own” — edits files without asking but runs terminal programs with your consent; “Full access” — no questions within the project. #### You see an edit before it happens A permission request shows the edit itself, like a diff. If you disagree, write what to do instead: the agent does it your way and carries on. #### Under every turn — what changed “2 files changed”: the list of files, lines and a diff. The agent’s work does not get lost in a long conversation. #### File edits roll back with one button Before a rollback you see what will be restored, with a warning if you edited the file after the agent. Changed your mind — “Restore as it was”. A rollback does not undo what programs it ran did outside the project: in “Ask first” mode such steps run only with your consent. [A permission request or a question from the agent also reaches your phone: connect it under “Settings → Phone” or with the `/phone` command — and answer without going back to your computer. Mobile app](https://buff.systems/en/mobile) **Permission request** “Ask first” mode ← scroll the frame → **The agent wants to add a line to the README:** the edit is visible before you allow it — and “Deny” with an explanation sends the agent another way. Objection 02 ## “The agent tackles a large task any old way” It starts editing before it understands what is wanted. ### First a plan, then the work — and a check of the result For a task across several files, turn on “Plan”: the agent first studies the code, asks where a decision is missing and brings a plan. It starts working after your “Approve and start” — and ticks off what is done in the list of steps. If the result can be checked, set a goal: the agent keeps taking turns until the check confirms it. #### A plan — before the first edit The agent studies the code, asks where a decision is missing and brings a plan. Work starts after your “Approve and start”. #### A list of steps — progress you can see Above the input field is a list of steps: how many are done out of how many, and what the agent is doing right now. #### A goal — up to a verifiable result For example, “all checks pass” or “the report is built and has a monthly total”: the agent works turn by turn, and a separate check after each turn says whether the goal has been reached. #### A side question — without derailing the task Ask while it works, in a tab alongside: the side conversation sees the whole session, while the main task does not see it. **Plan** awaiting approval ← scroll the frame → **The “how much is left to free delivery” plan:** the agent studied the code, asked questions where a decision was missing — and only then brought the plan for approval. Objection 03 ## “We run the same checks by hand every week” Run the tests, compile a summary of the week’s code changes, see what broke. ### Automations: a scheduled task for the agent Describe the task once and pick a schedule — daily, on certain weekdays, monthly. Buff Code runs it on its own, even with the window closed (the computer must be on), and puts the result into the project as a new session. How much the agent may do in such a run is set separately; if it needs your permission, the run waits for your answer. **Automations** two scheduled tasks ← scroll the frame → **A morning test check and a weekly summary:** you see when the next run is and how the last one ended; any of them can be run now. Objection 04 ## “Code review is a bottleneck: one person, one pair of eyes” There is no second reviewer for every change, and the first pass eats time. ### Workflows: several agents split the work — and combine the result Describe it in words: “two look at the file, each with its own job: one looks for calculation errors, the other for unclear places; the third combines their remarks into a summary.” The agent builds a workflow from this, checks it and runs it. You see which agent is doing what and what it costs, and at the end — the summary as a separate document. The agents do the first pass — the decision stays with a person. A good workflow is saved and run again — with another file. **Review by workflow** summary ← scroll the frame → **Review with three agents:** two agents looked at the file, each with its own job, and the third combined their remarks — from important to minor, with links to lines. Objection 05 ## “The agent does not know how we do things” It writes good code — but not yours. ### The company’s knowledge, trackers and tools — right in the work The connector is a small program your IT installs in your network; through it Buff sees your systems. The agent in Buff Code reads everything connected to it on its own: the release policy, a task in the tracker, a client contract, a neighbouring repository. And what only your team needs is connected in the settings. #### Tracker, reference data, documents Jira, GitLab, Confluence, Notion, Google Drive, folders on a server — everything connected to the connector. #### The organisation’s knowledge base The team’s rules, decisions and pitfalls — from the first message. The agent proposes new entries itself, but an entry appears only if you add it. #### Code of the whole project A neighbouring service, a shared library, settings in another repository — the agent finds and reads them without downloading them to you. #### Your own tools What only your team uses — internal services, your own ways of working, project notes — is connected in the settings, with no manual file editing. Anything sent from someone else’s repository will not run without your approval. **A knowledge base entry** awaiting your decision ← scroll the frame → **The agent itself proposed recording a team rule:** the entry appears in the knowledge base only if you add it. Objection 06 ## “Each developer has their own model subscription. Why change anything?” It is convenient until bills arrive from five different services and nobody knows which code went to which service. ### One model and one bill for the whole team The model is chosen by the organisation administrator — the same as in the rest of Buff. A developer needs no key of their own. Usage is charged in roubles on the organisation’s plan: the balance, how many days it will last and the approximate cost of the current session are shown right in Buff Code, and the spending history shows how much each device spent. | | Now | With Buff Code | | --- | --- | --- | | Who pays | Each developer, with their own subscription, then by receipts | The organisation, in roubles, on one plan | | Where the keys are | In the settings on every laptop | With Buff; there are no keys on the developer’s machine | | Which model | Everyone has their own | The one the administrator chose — one for the team | | Is spending visible | Only in each service’s personal account | The balance and your own spending — in Buff Code itself; by device — in the spending history | Objection 07 ## “Our systems are inside the network. Nobody will open them to the outside” And rightly so. ### Buff Code talks only to your company’s connector The connector is a program in your network through which Buff works with your code and systems. For Buff Code, a separate mode is turned on in the connector — the developer gateway: the team’s machines talk to it, and it talks to Buff and to your systems. - The developer’s machine has no direct connection to Buff — only to the connector inside your network - Passwords and keys for systems connected through the connector stay in the connector; the model does not see them - What goes to Buff is the sign-in confirmation and the conversation with the agent: the developer’s messages and what the agent read for its answer, such as project files. Buff stores the session record and passes the conversation to the model: on OpenRouter plans — through OpenRouter to third-party model providers outside Russia, with Buff’s key; on dedicated capacity and in a custom setup — to a model on a dedicated machine, and nothing goes to third-party providers - Buff stores a copy of the code — only of the repositories (code stores) the organisation itself connected to the project in the workspace. Buff fetches it from your code server (git): directly, with the access the organisation itself grants when connecting, or, if the server is closed to the outside, through the connector, where the access stays. Buff uses this copy for development tasks, and the agent in Buff Code reads from it the project’s other repositories that are not on the developer’s machine - The connection is secured by your company’s certificate. Buff Code adds no certificates to the system and does not change its network settings; the terminal program installs without administrator rights - Sign-in is by a code the developer confirms in the workspace (the Buff web app); a device is revoked with one button [The developer gateway in the documentation](https://buff.systems/en/docs/buff-code/gateway) Objection 08 ## “What remains of a developer’s work with the agent?” Security will ask, and so will the head. ### Every turn is recorded in full — in Buff, in your organisation’s data The request to the model and the response are stored in full — for your sessions, automations and workflows alike. The records are kept on the same terms as the organisation’s other materials and are visible only to it: a developer sees their own sessions, the organisation owner sees all of them, with totals by period, project and person. You can ask the model about the records — what the team investigated this week, where a newcomer got stuck. Every view of someone else’s session and every export goes into the audit log. [What exactly is recorded](https://buff.systems/en/docs/buff-code/sessions) **Development sessions** in the workspace ← scroll the frame → **Each session is labelled with the developer’s first question:** you see who, in which project, from which device, how many turns and what it cost. At the top — the Buff Code total for the period; bottom left — the balance and spending of the whole organisation. Objection 09 ## “How long will setup take?” A small team has no dedicated administrator. ### If the connector is already installed — one technical step Buff Code mode is turned on in the connector once — by whoever installed it. A developer needs only one command and a sign-in code; the organisation and project are chosen on first launch. *01* #### Whoever installed the connector turns on Buff Code mode in it — once A few lines in the connector settings and a certificate for its address in your network — from your internal certificate authority or an ordinary public one. *02* #### The developer installs Buff Code — with one command An app for Linux, macOS and Windows, or a terminal program. On first launch there is a short introduction: what you do, the theme, what to turn on. *03* #### Sign-in with a code from the workspace The connector address, a code on screen, confirmation in the workspace — and a choice of project. [No connector yet? Start with it](https://buff.systems/en/docs/connector/install) The terminal program — paste into a terminal (Linux and macOS; on Windows, in Git Bash). The app — on the [“Download”](https://buff.systems/en/downloads#buff-code) page. ``` curl -fsSL https://buff.systems/install | bash ``` Other questions ## What development teams ask ### How is Buff Code different from working in the Buff web app? In the web app, Buff runs the task: it writes the code itself, checks it and brings the changes to review. In Buff Code, an engineer works with the agent — in their own folder, on their own machine. The models, knowledge and connections are the same. ### A window or a terminal? It is one program in two forms: an app with a window and the buff program in the terminal. Sessions, automations, workflows and settings are shared; open a session in the window and continue it in the terminal. ### Can the agent break something in the project? You decide with the mode how much it may do without asking. Every turn leaves a summary of changes, and file edits roll back with one button; a rollback will not silently overwrite your edits made after the turn — it warns you. A rollback does not undo what programs it ran did outside the project — which is why in “Ask first” mode every run needs your consent. ### Does an automation run when the laptop is closed? It is started by the Buff Code service on that machine, so the window can be closed, but the machine must be on. Each run is a new session; its result waits for you in the list of sessions. ### Is a workflow programming? No: the agent writes and edits the workflow, and you describe in words who does what. A good workflow can be saved and run again — with another file. ### Where are session records stored? In Buff, in your organisation’s data, on the same terms as its other materials. A developer sees their own sessions, the organisation owner sees all of them; the records stay with the organisation after an employee leaves. ### What if an employee leaves? Revoke their device in the workspace: Buff stops answering them at once, and the connector closes access to your systems within a minute. ### Can I work on my own model subscription? No: Buff Code uses only the organisation’s models. That way the team has one model, one bill and one work history, and provider keys are not scattered across laptops. ### What language is the interface in? Russian by default; English in the settings. The language is the same in the window and the terminal and changes immediately. ### How much is it? There is nothing extra to pay: work in Buff Code is charged to the organisation’s balance on its plan, like any work in Buff. The balance and the approximate cost of a session in roubles are shown right in Buff Code. Where to start ## Install Buff Code and give the agent its first task in your project You need the company connector with Buff Code mode turned on. If there is no connector yet, start with it: the whole team needs it, not only for Buff Code. [Download Buff Code](https://buff.systems/en/downloads#buff-code) [Write to us](mailto:info@buff.systems) Companies and sole traders can pay by bank transfer and receive closing documents