Affordable to run all day
Point Tanya at DeepSeek and keep it running — routine work gets done without a frontier-model bill.
Public beta · v0.7.0 · free forever
Open-source CLI · built for DeepSeek
Real coding agent. DeepSeek prices. Verified output.
Tanya brings Claude Code-style flow to DeepSeek — then checks the final state itself, so a cheap model's confidence never becomes your 2 a.m. bug. Free, open-source, and it stays that way.
tanya — live sessionlive
▌ run_shell · xcodebuild -scheme CalculatorCompiling 6 files…** BUILD SUCCEEDED **▌ verify · final-state check✓ build passes✓ keypad wired · decimals correct✓ verdict: DONE — not "the model said so"
Free forever · open-source
Built in the open. Free today, free always.
Tanya is open-source on GitHub and there's no paid tier waiting in the wings. Read every line, file an issue, or ship a pull request — the project gets better because builders like you push it forward.
$0, and it stays $0
No subscription, no seats, no "pro" wall later. The only thing you ever pay for is your own DeepSeek tokens.
Fork it, bend it, ship it
The full source lives at github.com/matheusjkweber/tanya under an open license. Make it yours.
Contributions welcome
Bugs, ideas, skill-packs, docs — open an issue or a PR. Every contributor shapes where Tanya goes next.
Why Tanya is different
Cheap models hallucinate. Tanya checks their work.
Every other budget-friendly agent ships whatever the model claims it did. Tanya runs a deterministic verifier on the final state before it tells you the task is done.
A verdict you can trust
Per-task validators plus a final-state verifier mean "done" means done — not "the model said so."
Routes the hard steps
Tanya keeps deepseek-chat on the grunt work and escalates the genuinely tricky steps to deepseek-reasoner — automatically.
What's inside
Claude Code ergonomics, tuned for DeepSeek
Everything you expect from a serious terminal agent — streaming edits, sub-agents, MCP, permissions — tuned for a cheap model that needs a second pair of eyes.
The moat
Deterministic final-state verifier
Tanya re-checks the repo after it works and prints a real verdict. No more trusting a cheap model's word that the build passes.
tanya — live sessionlive
▌ run_shell · xcodebuild -scheme CalculatorCompiling 6 files…** BUILD SUCCEEDED **▌ verify · final-state check✓ build passes✓ keypad wired · decimals correct✓ verdict: DONE — not "the model said so"
Bring your own key
Built for DeepSeek
Tanya talks to DeepSeek over its OpenAI-compatible endpoint. Drop in your DeepSeek key and start coding — no glue, no wrappers.
Live by default
Watch every edit as it happens
Streaming tool execution shows files changing and commands running in real time — never a silent black box.
Spend where it counts
deepseek-chat ↔ reasoner routing
Keep deepseek-chat on the grunt work and route the genuinely hard steps to deepseek-reasoner, automatically.
Divide and conquer
Sub-agent task tool
Hand big jobs to focused sub-agents that plan, dispatch and report back — with optional TDD and auto-fix loops.
Connected
MCP server support
Plug in Model Context Protocol servers to give Tanya your tools, data and APIs — the same ecosystem you already use.
Stay in control
Permissions & spend rules
Approve risky actions, cap token spend and set guardrails so an over-eager run never surprises you.
Gets smarter
Golden-task memory & repo map
A structural repo map and golden-task memory help Tanya find its footing fast in code it has seen before.
Watch it work
Empty folder to working app, one step at a time
Tap through a real Tanya run: connect to DeepSeek, plan, edit files live, then verify the build before it ever says "done."
tanya — live sessionlive
$ tanya→ connecting to deepseek-chat✓ connected · OpenAI-compatible endpoint→ indexing repo · 218 files mapped✓ ready · type a task or /help$ tanya is listening_
Drop in your DeepSeek key and Tanya is live in one command.
The actual recordingreal run · sped up 2×
00:09
Tanya building and verifying a calculator app, end to end.
Roadmap
Where Tanya has been — and where it's going
Tanya already does the hard part: a real coding agent that works on cheap models. Here's the path so far and what we're building next, toward Claude-Code-grade power on your own keys.
ShippedIn progressPlannedLater
- Shipped
A coding agent that survives DeepSeek
Permissive tool-call parsing, retry-with-correction, schema flattening and reasoning-model support — the groundwork that makes cheap models usable.
Tool-call recoveryStreaming REPLReasoning models - Shipped
The verifier moat
A deterministic verifier, platform validators and a forbidden-pattern gate catch the hallucinations and shipping bugs cheap models would otherwise sneak past.
VerifierApple / Android / Go / Prisma validatorsForbidden-pattern gate - Shipped
Spend less, route smarter
Per-step multi-provider routing with a token-fit cost cascade, plus a live token + USD counter so the bill never surprises you.
Multi-model routingCost cascadeLive cost counter - Shipped
An agent you can extend
MCP client and server, slash commands, skill packs, project memory and bounded sub-agents — with session resume so long work survives.
MCP client + serverSlash commandsSkillsSub-agentsSession resume - You are here
- In progress
Extensibility core
The features that turn Tanya from an agent into a platform you can shape to your own workflow.
Lifecycle hooksNamed sub-agentsSlash-command templating - Planned
Daily-feel parity
The interactions that make a coding agent feel effortless every day.
Interactive plan modeHierarchical memory + importsWeb toolsBackground shells - Planned
Power-user & trust
Deeper control and safety for people who live in the terminal.
Checkpoint / rewindUnified settingsContext view + statuslineOutput styles - Later
Ecosystem
Reach beyond the CLI once the core is rock-solid.
Git / PR automationOS sandboxImage inputSDK
Have something we should build?
The roadmap is shaped by what builders actually need. Tell us what would make Tanya better — it lands straight in our inbox.
Pricing
Free, and open-source forever
Tanya itself costs nothing and always will. The only thing you pay for is your own DeepSeek tokens — and keeping that bill tiny is the whole point.
Tanya CLI
Free forever$0open-source · no account · free forever
- The full agent — verifier, routing, sub-agents, MCP
- Built for DeepSeek — drop in your API key and go
- Slash commands, permissions and spend caps
- You only ever pay DeepSeek for the tokens you use
No subscription, no seats, no telemetry paywall. Install from npm and go.
Bring your own key
Tanya is free. You just bring a DeepSeek key.
Tanya never touches your card — you pay DeepSeek directly for the tokens you use. The good part: DeepSeek is one of the cheapest models out there, with coding and reasoning scores in the same league as the big frontier models. Top-tier output, pocket-change bills.
Get running in three steps
1
Create a DeepSeek accountSign up at platform.deepseek.com — a minute, an email, done.
2
Mint an API key & add creditOpen the API keys page, generate a key, and top up a few dollars — that lasts a long way.
3
Hand it to TanyaSet DEEPSEEK_API_KEY (or paste it on first run) and you're coding.
Why DeepSeek?
Frontier-class benchmarks at a fraction of the token price — exactly the cheap-but-capable model Tanya's verifier was built to keep honest. You get near-frontier coding without the frontier invoice.
- typical cost per task
- ~$0.02 typical cost per task
- drop-in endpoint
- OpenAI-compatible drop-in endpoint
- auto-routed for you
- chat + reasoner auto-routed for you
Questions
The honest FAQ
Tanya is a public beta. Here's exactly what that means.
Is Tanya really free?
Yes — free today and free forever. Tanya is open-source and costs nothing to install or run, with no paid tier planned. You only pay DeepSeek for the tokens you use, and because Tanya is built around a cheap model, that bill stays small.
Which model does Tanya use?
DeepSeek. Tanya connects to DeepSeek's OpenAI-compatible endpoint — you supply your own DeepSeek key and run. It uses deepseek-chat for everyday work and can escalate the hardest steps to deepseek-reasoner automatically.
What does the verifier actually do?
After Tanya finishes a task, a deterministic verifier inspects the final state of your repo and per-task validators check the specific goal. You get a real pass/fail verdict instead of trusting that the model did what it claimed.
Where's the source code?
On GitHub at matheusjkweber/tanya, published to npm as @matheuskrumenauer/tanya. Issues and skill-pack contributions are welcome.
What do I need to run it?
Node.js and a terminal on macOS, Linux or Windows. Install globally with npm, set your DeepSeek API key, and run tanya. A hosted web playground is on the way.
How is this different from Claude Code?
Same ergonomics — live chat, slash commands, sub-agents, MCP — but built for DeepSeek, with a deterministic verifier as the headline feature because a cheaper model needs that check more, not less. And it's free and open-source.
Spin up Tanya in one command
Open a terminal, install from npm, point it at DeepSeek. That's the whole setup — and it's free.