One line to install.
380+ models as agents.
Sente is a thin launcher over OpenCode (MIT, 203k GitHub stars). One curl and every teai.io model is an agent in your terminal. Zero conversion layer — nothing to break when routing through teai.
curl -fsSL https://teai.io/te | sh
te # start an agent in the current dir
te run "explain this repo" # one-shot
te max run "..." # max perf = Kimi K3 (2.8T, 1M)
te -m teai/
te / sente / fuseki — call it, talk to it, let it work
teai.io's agent CLIs come in three stages, each with a different way of showing up.
| te | sente | fuseki | |
|---|---|---|---|
| Trigger | You type | You speak | Nothing — it's already watching |
| How it moves | Call it, it moves | Speak, and it moves first | Without being called, it lays the groundwork |
| Input | Text | Voice | Repo state, human-gates, etc. |
| Safety model | Run on request | 3-tier risk (read = auto / write = reversible, proceeds / delete·send·publish·pay = asks first) | Inherits the same tiers. Defaults to proposing only — never executes on its own |
| Status | GA | GA | Alpha · in design |
fuseki now has an Alpha implementation: fuseki / te watch. It keeps an eye on your board (human-gates, recent repos) without being called, and only thinks and logs+speaks a suggestion when something actually changes. It never executes anything on its own. Stop it with te stop. It inherits the same safety model te and sente already run in production (anything risky always asks first).
Why Sente
We compared 13 of today's coding agents. For use with teai.io, Sente is the best fit — here's why we're confident.
Zero conversion, stable by design
teai.io is natively OpenAI-compatible. Sente (OpenCode-based) speaks OpenAI-compatible natively, so tool round-trips never break through teai. Claude Code (Anthropic tool_use conversion) and Codex (proprietary format) can break depending on the model.
Inherit every upstream improvement
Not a fork — a thin launcher. On each launch it syncs the teai.io model catalog (380+) via /te/config, so you inherit every improvement to OpenCode (MIT, 203k stars) upstream.
Tokyo region, JPY billing
Tokyo region, JPY billing, Japanese support. Coding discipline (sente-rules.md) is injected into every session — "read before you write" and "always ship a deliverable" enforced mechanically.
One line to install
Just curl -fsSL https://teai.io/te | sh. Installs OpenCode if missing, points OPENCODE_CONFIG at the teai.io-generated config. te claude routes Claude Code through teai too.
380+ models, one line to switch
Daily driver: glm-5.2 (~¥0.34/task). Quality-critical work: te lux for the Claude/OpenAI flagship (Fable 5). Hard tasks: te max for Kimi K3 (2.8T, 1M). DeepSeek V4 Pro is ~¥0.03/task. One base_url decides routing — cheap without breaking quality.
1 task ≈ 1K in + 500 out. See bake-off for measured numbers.
Privacy & enterprise-ready
API-compatible endpoints never store request bodies — only metadata, kept 90 days. Invoice billing, DPA available. BYOK (bring your own key) is in preparation. See security & privacy.
Local PII scrubbing (opt-in, New)
te privacy scrub on runs a local Ollama model (qwen3.5:4b) plus regex on this Mac right before sending, redacting names, addresses, phone numbers and API keys into placeholders before the request reaches teai.io — the reply is restored before it's shown to you. If Ollama isn't reachable it fails closed (nothing is sent unredacted) rather than falling back silently. Not a guarantee of complete detection. Read more.
5 beta skills, backed by reference-corpus RAG
748 Q&A entries across law, security, freelance ops, cloud infra, and OSS licensing, searched with lightweight retrieval (semantic embedding + relevance cutoff) before answering. Full benchmark numbers — including where it failed — are published here.
te legal — Japanese law (259 entries)
Civil code, labor law, company law, inheritance, consumer contracts, commercial transactions — 21 topics. Cross-checked against actual e-Gov statute text; semantic search returns "no match" for unrelated questions instead of fabricating an answer.
te security — secure coding (131 entries)
SQLi/XSS/CSRF mitigation, auth/authz, secrets management, dependency vulnerabilities, crypto basics, API security — 13 topics. Sanity-check "is this dangerous?" mid-implementation.
te freelance — freelance ops (125 entries)
Contract checkpoints, Japan's invoice system, tax filing, social insurance switching — 14 topics. The non-coding stuff solo developers get stuck on.
te infra — cloud ops (117 entries)
Fly.io, Docker, CI/CD, SQLite/libsql, DNS, TLS — 14 topics, including real gotchas teai.io itself hit running this exact stack.
te license — OSS licensing (116 entries)
MIT/Apache/GPL-family, AGPL's SaaS network clause, license compatibility — 14 topics, written after checking each license's official text.
Callable directly from the API — shitate/legal
Beyond the te legal "question" CLI, pass {"model":"shitate/legal"} to /v1/chat/completions directly and the same RAG context gets spliced in automatically. Same alias-resolution pattern as teai/auto — billing reflects whichever model actually served the response.
te security "how do I prevent SQL injection?"
te freelance "how do I register for the invoice system?"
te infra "how do I set a secret on Fly.io?"
te license "what should I watch for using AGPL in a SaaS?"
All [β] — general reference information, not a substitute for professional advice. Full methodology and numbers in the benchmark post.
Sente in 15 seconds
Recorded from a real te session: install → run → coding demo.
te run → file generation task → switch to te max (Kimi K3)13 coding agents compared
As of Sep 2026. Compared for use with teai.io (Anthropic + OpenAI dual-compatible, Tokyo region, JPY billing).
| Agent | Form | OSS | Any LLM | Anthropic-compat | OpenAI-compat | Japanese | teai.io fit |
|---|---|---|---|---|---|---|---|
| Sente (te) teai.io official |
CLI | OpenCode (MIT) thin launcher | teai only, 380+ | via teai | native | native | best (zero conv) |
| Claude Code | CLI/IDE/Desktop | proprietary | Anthropic-centric | native | via teai | no UI | 2 lines to plug in |
| OpenCode | CLI/IDE/Desktop | MIT, 203k★ | 75+ providers | yes | yes | UI yes | Sente's core |
| Cursor | IDE/CLI/Cloud | proprietary | own + major | × | via OpenAI-compat | UI yes | official conn: own |
| Aider | CLI | Apache, 49k★ | almost all | via litellm | yes | unofficial only | via OpenAI-compat |
| Cline | IDE/CLI/SDK | Apache, 67k★ | all major + Ollama | yes | + any compat | UI yes | via OpenAI-compat |
| Gemini CLI | CLI | Apache, 107k★ | Gemini only | × | × | English only | incompatible |
| Codex CLI | CLI/Desktop/Web | Apache, 121k★ | OpenAI-centric | × | own | English only | own fmt (partial break) |
| GitHub Copilot | CLI/IDE/Web | proprietary | own routing | × | × | UI yes | no direct conn |
| Warp | Desktop | partial OSS | own + any | × | × | UI yes | AI terminal-ish |
| Cody (Sourcegraph) | IDE/CLI | proprietary | own + any | × | yes | English only | direction unclear |
| Tabnine | IDE/CLI | proprietary | multi-select | × | yes | English only | enterprise-ish |
| Continue | VS Code/JetBrains | ex-OSS | any | yes | yes | English only | acquired by Cursor |
★ = GitHub stars (as of Sep 2026). Sente is a thin launcher over OpenCode (MIT, 203k★), not a fork.
※ This table is on the teai.io official site. For fairness, full benchmarks are in the Sente benchmark report (note: Codex's breakage was a teai-side bug, since fixed).
teai.io exposes both /v1/messages (Anthropic-compat) and /v1/chat/completions (OpenAI-compat), so Claude Code plugs in via 2 ANTHROPIC_BASE_URL lines, and Sente/OpenCode/Cline/Aider via 1 base_url line.
For cost/quality benchmarks: cost/quality bake-off.
Structural conclusion: teai is OpenAI-compatible, Sente is zero-conversion
Same task, three clients, all routed through the teai backend.
| Client | Behavior via teai | Result |
|---|---|---|
| Sente (OpenCode) | Sends OpenAI-compat as-is. Zero conversion | ran clean to completion |
| Claude Code | teai converts Anthropic tool_use | model-dependent |
| Codex | proprietary function-call + sandbox | own fmt, partial break |
Start in 3 minutes
Paste one line in your terminal. Installs OpenCode if missing, points it at teai.io's config.
curl -fsSL https://teai.io/te | sh
# 2. API key (get a free one at /register — 100 credits on signup)
te login
# 3. run
te # interactive mode
te run "refactor this function"
te max run "..." # hard tasks on Kimi K3
te claude # route Claude Code via teai too
# 4. optional: GUI apps (macOS · Developer ID signed + Apple notarized)
te app install sente # menu-bar Sente.app
te app install koe # always-listening Koe.app
te app install both # install both
Works on macOS, Linux, and WSL (Windows: use WSL). Coding discipline (sente-rules.md) is auto-written to ~/.config/teai/ and loaded into every session. GUI apps are placed in /Applications only when you explicitly run te app install (transparency first). See the docs and the benchmark report.
Frequently asked questions
What is Sente based on?+
A thin launcher over OpenCode (opencode.ai, MIT, 203k GitHub stars) — not a fork. On each launch it syncs the teai.io model catalog (380+) and points OPENCODE_CONFIG at a teai.io-generated config, so you inherit every upstream OpenCode improvement.
How is it different from Claude Code or Codex?+
teai.io is natively an OpenAI-compatible gateway. Sente (OpenCode-based) speaks OpenAI-compatible natively, so there's zero conversion layer and nothing to break when routing through teai. Claude Code relies on teai converting Anthropic tool_use, and Codex uses its own function-call format plus a sandbox — both can break on certain models. Details in the benchmark report.
What does "Sente" mean?+
teai.io is "手AI" (hand + AI). Sente inherits the "hand" and is named after the Go term sente (先手, initiative/first-move). The agent moves before you do. "Always keep sente" — never let go of the initiative. The command is short: te (a sente alias is also installed).
What does it cost?+
Sente itself is free. teai.io uses a credit system: the Free plan gives 100 credits on signup (no credit card required) — enough for ~300,000 short chats on Qwen3.7 Flash or ~2,000 on glm-5.2 (1K in + 500 out each). Pro (¥4,350/mo ≈ $29 USD, 30,000 credits) and Business (¥14,800/mo ≈ $99 USD, 100,000 credits) are also available. See the pricing page.
Try it free now
100 credits on signup. No credit card. Start in 30 seconds.