/new · /chats
Parallel conversations
Every chat owns its own forked session and agent task. Background turns keep streaming while you type somewhere else — nothing ever collides.
── installation
One file on Windows. One line everywhere else.
Prebuilt for five targets, an installer that never asks for admin rights, and a shell script that checks SHA-256 before it puts anything on your PATH.
⤓ pooprusteek-setup.exe
- 1.Run pooprusteek-setup.exe. It installs per user into %LOCALAPPDATA%\Programs, so Windows never asks for admin rights.
- 2.One screen: pick the folder, press Install. It adds pooprusteek to your PATH and a Start-menu shortcut that opens in Windows Terminal.
- 3.Uninstalling is an entry in Apps — and it asks whether your settings and sessions should go too.
! The installer isn’t code-signed yet, so SmartScreen may warn on first run: More info → Run anyway. Every file’s SHA-256 ships with the release if you’d rather trust math than a vendor.
without the wizard
Pull the installer from PowerShell and run it — same file, no browser involved.
Or skip installing altogether: the portable archive carries the same binary — pooprusteek-windows-x86_64.zip · arm64
It updates itself: /update · /autoupdate on
── switching over
Coming from Claude Code, Codex or Gemini CLI?
Same job — an agent in your terminal that reads code, runs commands and edits files — without the subscription. And your project setup comes with you.
── capabilities
Small binary. Senior engineer.
«Работает как старший инженер: автономно, хирургически, без воды.» — the system prompt, and it means it.
/agent · task
Sub-agents
Spawn isolated workers — foreground for a clean answer, background for fire-and-forget. Only the conclusion comes back, not the noise.
/goal
GOAL loop
A worker writes, an evaluator judges. The loop iterates on concrete feedback until the goal passes, and swaps sessions to escape a dead end.
/rag · /search
Local RAG
Multilingual embeddings and keyword search over your skills, MCP tools and every past conversation. Your CPU, offline, after one 120 MB download.
/mcp
MCP, with OAuth
stdio, HTTP and SSE servers, auto-discovered from eight config sources. Full RFC OAuth, with tokens in the OS keyring instead of a plaintext file.
/skills
Markdown skills
Reusable instruction sets found across a dozen agent directories. Pin one by hand, or let the matcher offer it the moment it applies.
/providers · /models
Bring any model
DeepSeek web by default; OpenAI-compatible, Anthropic and Gemini endpoints alongside it. Switch from a panel, pick models from the live list.
/serve
Local API gateway
One command turns the agent into an OpenAI-compatible server on 127.0.0.1. Every tool that speaks that dialect inherits the free backend.
edit · write · /undo
Surgical edits
Anchored replacements with a real diff preview, not sed. Every write is checkpointed first, so /undo puts the file back the way it was.
bash · /jobs · /ps
Real shell, real PTY
Foreground, background and interactive PTY processes. Dev servers survive turns on an idle TTL; /jobs and /ps list and kill them.
/whitelist
Approvals with a scope
Allow cargo test, not all of bash. A rule is a tool plus a scope — and a compound command gets none, because it only looks like its first word.
/update · /themes
Yours to keep
Self-update on two channels, verified against the release manifest. Ten built-in palettes, plus a wizard for one of your own.
monthly bill
$0.00
Not a trial. Not a free tier with a token ration. PoopRusteek drives the same web chat you already use for free — it just does it from your terminal, with tools.
pow sha3(nonce=48213) → 0003dc90b… solved locally
- auth
- your own chat.deepseek.com session — cookie + userToken, stored in the OS keyring, encrypted at rest
- proof-of-work
- the SHA-3 challenge DeepSeek uses to gate its web API is solved locally, on your CPU
- api key
- none. there is nothing to leak, rotate, or top up
- telemetry
- none. your code goes to the model you chose and nowhere else
── /goal mode
It doesn’t stop when it sounds done. It stops when it is.
Two agents, one loop: a worker does the job, an evaluator refuses to be impressed. You watch the status bar.
[GOAL ON]
worker + evaluator armed, waiting for a goal
── local rag
It finds its own skills, tools and memories.
A semantic layer that never leaves your machine: multilingual embeddings and stemmed keyword search, fused, over three corpora. One 120 MB model downloads once — after that nothing goes out.
every prompt, before it is sent
❯prompthow do I ship a release?
▚densee5-small · ONNX · 384-dim · quantized
▚lexicalSnowball TF-IDF · ru + en stems
⊕fusionreciprocal rank fusion
✓hintskill: releasing (0.93) → loaded on demand
Cross-language by construction: a Russian prompt matches an English skill description, and the other way round.
skills
Every outgoing prompt is matched against the whole catalog; the top hits ride along as an ephemeral hint. The model discovers what it can do instead of you enabling it by hand.
mcp tools
With thirty tools connected the prompt carries one line per server. Full schemas arrive only when a tool matches — or when the model asks for them with tool_search.
history
Every saved session is chunked, embedded and kept. Search it yourself with /search, or let the agent recall the fix it already wrote via history_search.
0.927
MRR over the skill catalog, from the repo’s own eval harness
0.836
MRR over MCP tool descriptions
120 MB
downloaded once — e5-small, quantized ONNX
0
bytes over the network after that
── the surface area ──
50 slash commands. Zero menus.
Everything is a command, exactly where your hands already are.
Every one of them is a file in src/commands/defs/, and the help screen is generated from the same registry. /architecture →
── /serve
Your free backend, on localhost.
One command turns the agent into an OpenAI-compatible server. Anything that speaks that dialect — aider, Continue, Open WebUI, the OpenAI SDKs — points at 127.0.0.1:7667 and inherits the same free session.
●127.0.0.1:7667 · openai dialect
❯ /serve on
API server listening on http://127.0.0.1:7667/v1
$ curl http://127.0.0.1:7667/v1/chat/completions \
-H 'content-type: application/json' \
-d '{"model":"deepseek-chat","stream":true,
"messages":[{"role":"user","content":"hi"}]}'
data: {"choices":[{"delta":{"content":"Hi"}…
$0.00one dialect, every backend
Model ids route to the built-in DeepSeek client or to any /providers entry: deepseek-chat, lmstudio/qwen2.5-coder, gemini/…
stateless per request
Every completion runs on a fresh fork, and DeepSeek forks drop their remote session afterwards — server traffic never piles junk chats onto your account.
loopback by default
127.0.0.1 unless you say otherwise. Set a bearer token and CORS turns on with it, not before.
No interface needed: /serve →
── /themes
Ten palettes. This page wears them too.
The gallery below is the agent’s real preset table from src/tui/theme.rs. Click one: it is what the TUI looks like, and it is what this site looks like — both are built on the same color roles.
Your own is a wizard away: /themes new takes a base preset plus per-role overrides and lands as a [[ui.custom_themes]] entry in config.toml.
1
static binary — no node, no electron, no runtime
62k
lines of Rust on tokio + ratatui
0
garbage-collector pauses while you type
5
prebuilt targets: Windows, Linux and macOS
Rust edition 2024 · MSRV 1.91 · ~900 tests on Windows, Linux and macOS · one tokio::select! loop, everything else is an event