A blistering-fast terminal coding agent that runs on your keys, your models, your machine. The complete agent stack — plan mode, sandboxing, subagents — in one tiny binary. Free with your own API keys.
Model providers (BYOK)
130+
Servers your code touches
0
Enforced sandboxing
OS-level
Single fast binary
~12MB
Bring your own key for
Anthropic
OpenAI
Google
Mistral
Groq
xAI
Ollama
+123 more
// SEE IT RUN
The agent works. You stay in control.
tachyon — ~/projects/api-gatewaysandbox: on · net: denied
add rate limiting to the /auth endpoints, then run the tests
Thought: 0.4s
Mapping the routes first — I'll stay in plan mode until the approach is approved, then apply edits and verify with the suite.
▣ read · src/routes/auth.ts · 118 lines
▣ grep · rateLimit · 0 matches
▣ read · package.json · 76 lines
◇ plan mode — mutating tools blocked while I map this out
1. add express-rate-limit to auth.ts — 100 req / 15 min per IP
auto-allowed — read-only test command, matches safe policy
✓ 48 passing (2 new) · 0 failing 1.9s
◍ done in 11s · /undo revert · /fork branch
❯
ask · deepseek:deepseek-v4-pro◆ think● max
Why Tachyon
01
Your keys. Your code. Your model choice.
Works with 130+ model providers via BYOK — you're never locked into one vendor's models or subscription. The tool is free with your own API keys, and your code never routes through our servers unless you opt in.
No lock-inFree with your keysLocal-first by default
02
Safe autonomy, enforced by the OS — not by promises.
Every session runs inside real OS-level sandboxing (bwrap on Linux, Seatbelt on macOS, Docker anywhere) with network access denied by default. Plan mode hard-blocks mutating tools instead of just asking nicely, and an AI permission layer auto-allows safe commands while flagging risky ones — speed without giving up control.
bwrap · Seatbelt · DockerNetwork denied by defaultPlan mode hard-block
03
The full agent stack in one fast binary.
Plan mode, hooks, skills, subagents, session fork / undo / redo, LSP-aware self-correcting edits, command palette and transcript search — the complete feature sweep only the biggest incumbent matches, in a tiny footprint. It even auto-imports your existing Claude Code and OpenCode MCP config, so switching takes seconds.