Driven from Telegram · runs on your machine

The coding agent that finishes the ticket.

AutoCoder runs on your Windows machine and is driven from Telegram. Send it a ticket and five agents — Discoverer, Planner, Executor, Verifier and Reviewer — work it end to end, narrating their reasoning into your chat, then hand back the diff and the exact token cost.

50 free
queries / month
BYOK
DeepSeek or Claude, your key
Notion
as long-term memory
AC
AutoCoder bot
online
⋮
/task add rate limiting to the /upload route
✓✓ 14:02
Message
the Telegram thread, verbatim — one message that keeps rewriting itself
01No AutoCoder server in between
02~5 min install · no card
03Drive it from Telegram
04Available on Microsoft Store →
Assign the ticketWalk awayRead the diffKnow what it costRuns on your machineYour key, no markup
Ch.1
Why this exists
Read this one first

What is AutoCoder — and how does it actually work?

The three things people want to know before they install. If you only read one section of this page, read this one.

01

What is AutoCoder?

AutoCoder is an autonomous coding agent that runs on your own Windows machine and is driven from Telegram. You assign a ticket from your phone and it works it end to end — reading the repo, planning, editing, running the tests and reviewing its own change — without waiting for you to approve each step. When it finishes, the diff and a written report arrive in your chat with the exact cost of the run.

02

How does it work, end to end?

You send /task with what you want done. Your desktop app is already polling for work, so it picks the task up and starts a five-stage pipeline: Discoverer, Planner, Executor, Verifier, Reviewer. As it works, its one-line reasoning streams back into the same Telegram message, rewritten in place, so you can watch progress without opening your laptop. It finishes by running your tests and handing back the diff.

03

What do I need to run it?

A Windows 10 or 11 machine that stays awake while a task runs, the free AutoCoder app from the Microsoft Store, and a DeepSeek or Claude API key — you pay that provider directly at their list price, with nothing marked up in between. A Notion key is optional and gives the agent long-term project memory between tasks. Install takes about five minutes; the first 50 queries each month are free.

The shift

Pair programmers need you in the chair. This one doesn't.

Cursor and Claude Code are excellent when you're at the keyboard, reviewing as you go. AutoCoder is for the other half of the work — the tickets you'd rather hand off entirely.

Agent in the loop
With AutoCoder
01 · in the loop
It only moves while you watch it
Approve a diff, approve a command, approve the next one. The agent is fast; the loop is you.
01 · with autocoder
It runs the whole task unattended
Five agents hand off to each other — discover, plan, execute, verify, review — with no confirmation prompts to babysit.
02 · in the loop
Progress stops when you leave the desk
A bug you thought of on the train waits until you're back at the keyboard to even get started.
02 · with autocoder
Start it from your phone, anywhere
Send /task in Telegram. Your desktop picks it up, narrates progress into the chat, and hands back the finished work.
03 · in the loop
Nights and weekends are dead time
The queue of small, well-specified chores never gets touched because nobody wants to spend Saturday on it.
03 · with autocoder
Queue it for 2am
The built-in scheduler runs tasks on a cron while you sleep, then reports back with the diff and what it cost.
04 · in the loop
You find out what it cost in the invoice
Subscription tiers and usage meters tell you how much is left, not what any given task actually consumed.
04 · with autocoder
Every task is itemised
Per-task and per-agent breakdowns in the app, /total_cost in Telegram, billed by your provider (DeepSeek or Claude) at their rates.
By the numbers

What it actually costs, and what it actually catches

Our own figures come straight out of the app — every number below is one /total_cost report away.

$0.038
a typical task
19 API calls, DeepSeek V4 Flash
95%
prompt cache hit rate
across 28 tasks — $3.98 saved
5
agents per task
or one, for simpler work
50
free queries / month
no card, no trial clock
task cost varies with task size and model · V4 Pro runs roughly 3× V4 Flash · nothing here is an estimate
Check it yourself

Everything above is checkable, and so is everything below

Microsoft Store
Listed under Developer tools, free to install
Distributed through the Microsoft Store — not an unsigned download from a landing page. Free to install, and it arrives signed.
Open the Store listing →
Your key, your bill
DeepSeek's list price, nothing added
A typical run costs low single-digit cents because there is no resale margin to add: the key is yours and you pay your model provider directly.
How the token spend is billed →
On the record
The claims, in the manual
Setup, the five-agent pipeline, the planner approval gate and the cost commands are all documented in the user guide.
Read the user guide →

We don't have customer quotes yet — so we're not going to show you any

AutoCoder is young. There is no wall of logos, no 5-star average, no “trusted by 10,000 engineers” banner, because none of that is true yet and you would find out in about forty seconds. What we have is a working product you can run on your own machine, on your own key, for free — and a cost report that tells you exactly what it spent.

Run it on one real ticket from your own backlog. If the diff is good, you already know more than any testimonial would have told you.

actively shippingWindows 10/11 · x64first 50 queries free, every monthpremium then ₹899/mo, tokens billed by your provider
Ch.2
What it does
Capabilities

Built to run without you

Everything you need to hand off a task and trust what comes back.

01 · remote
Drive it from Telegram
Send /task from your phone, watch live narration land in your chat, and get the finished work back the same way. Your desktop does the work — you don't have to be at it.
/task/cancel/total_cost/access_key
02 · autonomous
Delegate and walk away
Five specialised agents research, write, verify, and review in sequence — with no confirmation prompts to babysit. Drop to a single agent for simpler work.
03 · scheduled
Scheduled tasks
Queue work for tonight on a cron and let it run while you sleep. The diff and what it cost are waiting when you get back.
04 · local
Runs on your machine
The agent loop, your files, and project memory stay local, and your API key never leaves the app. No tunnel, no AutoCoder server in between.
05 · economics
Your key, no markup
Bring your own DeepSeek or Claude key and pay the API rate directly. We never proxy your requests or resell your tokens.
Know what every task cost
Per-task and per-agent breakdowns in the app, /total_cost in Telegram. No credits, no opaque meter.
Two models, one app
Use DeepSeek V4 (Flash or Pro) or Claude (Haiku or Opus) — switch per task, so cheap work stays cheap and hard work gets the stronger model.
Thinking budget
Cap how many tokens your model spends on reasoning, DeepSeek or Claude alike. More budget means deeper reasoning — and a bigger bill.
Local project memory
A cached repo map and findings store mean agents start each task already knowing your codebase.
Living rules & governance
Persistent project rules the agents re-read and follow on every single task.
Automated code review
Local Semgrep packs plus line-level findings with severity — JSON that CI can parse without glue code.
Long-term AI memory in Notion
Your team's conventions, decisions, glossary and known issues live in a Notion database and are injected into every task.
22 Indian languages
Write prompts with @regional and get research-backed answers back in the same language, at the English price.
How it works

Four steps, then it's out of your hands

Set-up is a one-time thing. After that the loop is: send a task, do something else, read the result.

Step 1 of 4install
/start
Account created — Free plan, 50 queries/month.
5X9A2C
Access key · rotates in 7 days
The pipeline

What runs inside step three

Five specialised agents, each with a narrow job, handing off in sequence. Simple tasks can run on a single agent instead.

cost-routed across V4 Flash / V4 Pro
01
Discoverer
Reads the codebase and researches the web — live docs, CVEs, current APIs.
02
Planner
Turns the ticket into an ordered plan and decides which files to touch.
03
Executor
Makes surgical edits rather than rewriting files — fewer tokens, less to re-verify.
04
Verifier
Runs typechecks, tests, and Semgrep locally, then loops back on failures.
05
Reviewer
Line-level findings with severity, ready for CI or your own eyes.
Automation

The queue that runs while you sleep

Hand the scheduler a cron expression and a task. It runs unattended, returns a grouped diff, and reports what it cost.

0
minute
2
hour
*
day
*
month
*
weekday
every night at 02:00 — the diff is waiting when you wake up
Recurring automation
Daily, weekly, or custom cron execution with zero manual intervention.
Telegram notifications
Completion reports and cost summaries land in your chat the moment a run finishes.
Grouped, reviewable diffs
Every run comes back as a diff grouped by intent, not a wall of changed lines — read it over coffee and decide what ships.
Review

Every change gets read before you read it

Each agent works from both sides at once — live web research for CVEs and current API docs, plus deep local analysis with git-diff precision.

Context-aware
Knows the latest external knowledge — CVEs, deprecations, current APIs — and the precise local codebase state.
Privacy-preserving
Local Semgrep scanning across three rule packs (secrets, OWASP Top 10, security-audit) with zero-egress project memory.
Resilient by design
Independent circuit breakers, TTL caching, and graceful degradation keep the pipeline running when an external service is down.
Cost-efficient
Cheapest-model routing for reviews, cache-first web search, and structural digests for context. No wasted tokens.
Structured output
Per-line annotations with severity, line numbers, and concrete fixes — consumable by CI and IDEs without a parser of your own.
CI/CD ready
Reviews fire automatically after every edit and emit JSON that GitHub Actions, GitLab CI, and Jenkins read natively.
Reach
@regional

Answers in your language

Prefix any prompt with @regional and get research-backed answers back in the same language — follow-ups included. Prompt in 22 Indian languages.

No premium tier, no surcharge — regional prompts cost exactly what English prompts cost.

Long-term memory

Notion is AutoCoder's long-term memory

Paste your Notion API key once. AutoCoder creates a small context database in your own workspace and injects your team's conventions, decisions, glossary and known issues into every single task — so it remembers between runs like a teammate who was there.

A
Conventions, decisions & glossary
Read from Notion and injected into every prompt.
B
Paste your key once
A small context database in your own workspace — no separate service to host, no data to migrate.
C
Save any answer
Drop a useful answer straight into the context database — it is there for the next task, and the one after that.
D
Shared team memory
Run history and the discovery cache live in Notion too, so the whole team inherits what the agent has learned.
Remote control

Everything you need, from a chat window

Account setup, key rotation, billing, and live task submission all happen in Telegram. There is no web dashboard to log into.

1 · Join the bot
Message @AutoCoderOfficialBot inside Telegram.
2 · Get your key
Run /start to register and receive a 6-digit access token.
3 · Log in
Enter the key in the desktop app and start assigning work immediately.
Command reference — pick one
AC
AutoCoder Official Bot
online
/task add rate limiting to the /upload route
Queued on your desktop
Narration streams here as it runs — no need to stay connected. Use /cancel to stop it.
Ch.3
Where it's the wrong tool
Comparison

Where it's different — and where it isn't

Rows we lose are in here on purpose — including the model row, where you should assume we lose. A table we won outright would just be marketing.

CapabilityAutoCoderCursorGitHub CopilotClaude Code
Control from a chat appTelegram — submit, steer, receive Only———
Scheduled / unattended runsBuilt-in scheduler—Via ActionsVia external cron
Runs without confirmationsDefault behaviour Hands-freePartialBasicConfigurable
Who your code is sent toYour model provider only, via your keyCursor servers, plus the providerGitHub / MicrosoftAnthropic only
Bring your own API keyRequired — you pay list price No markupOptionalNoOptional (API mode)
Multi-agent orchestration5 agents, or single-agent modeSingle agentSingle agentSub-agent delegation
Regional language support22 Indian languages, no surcharge Only———
Model capability ceilingDeepSeek V4 (Flash / Pro) and Claude (Haiku / Opus), your choice per taskFrontier modelsFrontier modelsClaude Opus / Sonnet Strongest
Editor integrationStandalone desktop appFull IDE DeepestDeep IDE integrationCLI + IDE extensions

To be precise about privacy: the agent loop, your files, and your project memory stay on your machine, and your API key never leaves it. Code an agent puts in a prompt still goes to the model provider you chose — DeepSeek or Anthropic — under your own key. There is no AutoCoder server in between, and we never see either one.

The honest answer

So which one should you reach for?

These are different tools for different moments, not rivals. Most developers who use AutoCoder keep a pair programmer open too.

Reach for Cursor / Claude Code
You want to review every edit as it lands
You want a single vendor's tooling and editor integration
You live in the editor and want the agent there too
You're exploring an unfamiliar architecture and want to steer
Reach for AutoCoder
The ticket is well-specified and you'd rather not watch it
You're away from the desk and want work moving anyway
It should run overnight, on a schedule, unattended
Your code and your API key must stay on your own machine
You want to pick the model per task: DeepSeek for cost, Claude for hard reasoning
Ch.4
What it costs
Pricing

A subscription plus your own token spend

Both halves, always stated together — the subscription alone isn't what running AutoCoder costs.

Free
Enough to find out whether the workflow suits you.
₹0 /month
+ your model token spend (DeepSeek or Claude)
50 queries per month
Bring your own DeepSeek or Claude key
Runs locally on your machine
Access to 2 specialised agents
Community support
Unlimited queries
All 5 agents in parallel
Start free on Telegram
Send /start — no card required
Premium
Most popular
Full autonomy — all five agents, no query ceiling.
₹899 /30 days
≈ $11 · + your model token spend (DeepSeek or Claude) · yearly saves 17%
Unlimited queries
Bring your own DeepSeek or Claude key
Runs locally on your machine
All 5 agents in parallel, or single-agent mode
Scheduled and recurring tasks
Premium Telegram community
Priority support
Go Premium
/buy_package — Stripe checkout

The other half: what DeepSeek and Claude charge you

Tokens are billed by DeepSeek or Anthropic (Claude), on your key, at their published rates — we never proxy your requests or add a markup. AutoCoder caches aggressively, so most input tokens bill at the cache-hit rate.

ModelInput / 1MOutput / 1MCached input
V4 Flash$0.14$0.28$0.0028
Claude Haiku 5.5$0.10$0.50$0.01
Claude Opus 5.5$4.00$20.00$0.20
Rates as published by DeepSeek and Anthropic. A typical task on V4 Flash lands in the low single-digit cents; Claude Haiku 5.5 bills a 5× surcharge on prompts over 100K tokens, and Opus 5.5 is flat-rate across context lengths. Both plans need the desktop app — download for Windows.

On macOS? Not yet.

The desktop app ships for Windows today, and we would rather say that plainly than let you find out at the Store page. Two things you can do instead.

Talk to the bot from anywhere
The Telegram side needs no install and works on every platform. You still need a Windows machine somewhere to run tasks against — but if you want to see how the workflow feels first, start here.
Open the bot
Get told when there's a build
No date to promise. One email when it exists, and nothing else.
FAQ

Questions worth asking first

Something not covered? Ask the bot directly or email support@autocoder.dev.

01Does my code get uploaded anywhere?
The agent loop, your files, and your project memory stay on your machine, and your API key never leaves the app. What an agent puts in a prompt does go to your chosen model provider (DeepSeek or Anthropic), under your own key — that's how any hosted model works. The difference is that there is no AutoCoder server in between, so we never see your code or your key.
02What will this actually cost me per month?
Two things, and you should count both: the subscription (₹0 on the free tier for 50 queries, ₹899/month for unlimited) plus your own model token spend (DeepSeek or Claude), billed directly by that provider at their rates. A typical task on DeepSeek V4 Flash lands in the low single-digit cents. Run /total_cost in Telegram at any time for the running total.
03Do I have to leave my computer on?
Yes — and awake. AutoCoder runs the orchestrator in-process on your desktop; Telegram is the remote control, not the runtime. Tasks you send from your phone queue on that machine and execute there, so if it sleeps or you shut the lid, the run pauses until it wakes. Before scheduling anything overnight, set the machine not to sleep. That's the honest trade for keeping your code and your key local — there is no cloud runtime to fall back on.
04Which platforms are supported?
The desktop app ships for Windows, via the Microsoft Store. On macOS there is no native build yet — that is the one thing on the macOS waitlist. On Linux, the CLI and the Telegram bot are the path today, and the bot works from any platform.
05Do I need my own API key?
Yes — it's required, not optional. Bring a DeepSeek key or a Claude (Anthropic) key, whichever you prefer, and pay that provider list price directly. We don't resell tokens, bundle credits, or take a cut, which is why there's no markup to explain.
06Why do I need a Notion key?
Because Notion is where AutoCoder keeps its long-term memory. Paste your Notion API key into the app once (Settings → Integrations) and AutoCoder creates a small context database inside your own workspace — your team's conventions, decisions, glossary, and known issues live there and are injected into every task. Like your model key, it stays on your machine: requests go directly to Notion under your own key, and nothing is proxied through an AutoCoder server.
07How does AutoCoder remember my project between tasks?
Before every task, AutoCoder pulls the relevant rows from your Notion context database — conventions, decisions, glossary, and known issues — and adds them to the prompt, so the agent starts each run already knowing how your team works and what it's decided. On top of that, it logs every finished task as run history and keeps a team-wide discovery cache of codebase findings, so knowledge compounds instead of resetting with each run. You can also save any individual answer to the context database from the app.
08What happens when an agent gets it wrong?
Two of the five agents exist for exactly that: the Verifier runs typechecks, tests, and local Semgrep packs and loops back on failures, and the Reviewer produces line-level findings with severity levels. You still get a grouped diff to read before anything ships. And /cancel stops a running task from your phone.
09Can I keep it from doing things I don't want?
Persistent project rules live alongside the repo and are re-read on every task, so conventions stick without repeating them in each prompt. You also control the thinking budget — how many tokens the model may spend reasoning — and whether a task runs on all five agents or just one.
10Does it fit into CI?
The reviewer emits structured JSON with per-line annotations, severities, and fix suggestions, which GitHub Actions, GitLab CI, and Jenkins parse natively. Scheduled tasks return the same grouped diff a chat-box run does, so you review it before anything ships.
11Why not just use Claude Code or Cursor?
Because they are built to keep you in the loop, and this is for when you don't want to be. That is the whole difference, and it is a difference of shape rather than quality — they are excellent pair programmers, on frontier models, and we are not trying to be one.

Underlying model: AutoCoder works with the model you choose, DeepSeek or Claude, on your own key. Claude Code and Cursor serve their own frontier models. If the next model release is what decides it for you, they are the honest answer, and we would say so on the comparison table rather than leave you to discover it.
12Is the 6-digit access key safe?
It rotates every 7 days automatically, and you can force a new one at any time with /access_key. It authenticates your desktop app to your own Telegram account — it is not your model key, which never leaves your machine.
13Can I write prompts in a language other than English?
Yes — prefix a prompt with @regional and you'll get research-backed answers in 22 Indian languages, at the same price as English. No surcharge, no separate tier.
From the blog

How we think about agentic coding

All posts →
AICoding Agent Harness Explained: The 7 Parts That Decide SafetyA coding agent is a model wrapped in a harness. The harness sets what the agent can read, run and change. Here are the seven parts that decide safety and cost, and how to check each one.Devdatta Narawade · 29 October 2026 · 9 minAIAI Coding Agent Security: Risks Your Approvals MissApproval answers whether a vendor is trustworthy, not what the agent can do with your credentials. Here is the real attack surface and a control ladder to act on this week.Devdatta Narawade · 28 October 2026 · 6 minAIAI Coding Agents Architecture: Why the Harness Matters More Than the ModelMost teams compare AI coding tools by model. In 2026 the harness decides whether an agent is safe, affordable and finishes work. Here is how the architecture fits together and how to audit yours.Devdatta Narawade · 27 October 2026 · 6 min
Start tonight

Give it a ticket tonight. Read the diff over coffee.

Fifty queries a month are free, your key stays on your machine, and the whole thing takes about five minutes to set up.

01Download the app
02Send /start to the bot
03Assign your first ticket