Outperform
open.
Rehex bets on the harness, not the model — so a $0–$0.30/M open model competes with a $15/M frontier one.
Frontier isn't free.
Thin harnesses wrap one expensive model and hope it's right — at real economic, environmental, and human cost, with someone always babysitting it. Rehex flips it: the harness does the heavy lifting, so a small open model carries the load for a fraction of the price.
Prune the context, keep the point.
Because Rehex is task-first, it can reason about how work gets completed as it's requested. Episodic compaction prunes large stretches of the context window while retaining what matters to the request — freeing tons of context and cutting round-trip tokens.
// green = retained signal tied to the request & the work
Pass a pointer, not the payload.
Drawing on recent recursive-language-model research, Rehex references large inputs by pointer instead of re-inlining them into every turn. Managing input tokens this aggressively has cut round-trip tokens by a meaningful margin — we're still measuring the average across sessions.
fewer round-trip tokens in some sessions
// peak observed — average still being measured
The toolkit
What the agent can do.
A broad, production-grade toolkit — and every tool is gated by DTS, so the agent only reaches for the right one at the right time.
File + code
Read, write, edit, and search files — glob, regex grep, line slices, multi-edit atomics, find/replace. Works on your codebase directly, no sandbox.
Shell + processes
Run commands, launch long-running background processes, tail logs, kill processes. 60s on quick commands; unbounded for overnight jobs.
Auto-git
Optional and nice — Rehex auto-commits as it works: init, status, commit, diff, branch through a typed interface, so every change is captured.
Web + research
Search the web, fetch and parse full pages as text, markdown, HTML, or JSON. Research, read docs, gather live data.
Browser automation
Full browser control — navigate, click, type, fill forms, scroll, screenshot, extract DOM. Drives a real browser like a human would.
Screenshots
Full-page or viewport captures at desktop, tablet, or mobile sizes — visual inspection without manual effort.
Vision + audio
Analyze images with vision AI; transcribe audio (mp3, wav, webm, ogg, flac) with speaker timestamps via local Whisper.
Knowledge base
Query internal docs with BM25 search, full reads, and AI-synthesized answers with citations. Plug in your own sources.
Subagents
Spawn independent sub-agents for parallel work — each with its own session, optional git worktree, and model override.
Task management
Create, list, complete, and delete tasks. Plan upfront, execute step by step, with the harness enforcing the lifecycle.
Session management
Open parallel sessions, close and merge work, compact context, manage worktrees. Orchestration for complex workflows.
Headless + TUI
Run headless in CI or scripts, or drop into a full TUI — and jump between them mid-session to inspect exactly what happened.
Isolation & accountability
Set it loose.
No permission prompts to babysit. Rehex runs under a real read / write / execute permission structure and keeps every action inside the directory it started in — working on your behalf seamlessly, with full git history and the ability to rewind whenever you need.
Scoped to one directory
Every read, write, and execution stays inside the directory Rehex started in. Nothing escapes the boundary.
R / W / X permissions
A genuine permission model — not a prompt asking you to approve each step. Set it once and let it run.
Git history, rewind anytime
Every action is committed and accountable. Roll back to any point the moment you need to.
Gated, not guessed.
Smaller models stumble on knowing which tool to use — a real barrier in production. DTS gates which tools the model can even see, evaluating declarative conditions against live session state before every turn. It can't skip steps, finish early, or grab what's irrelevant.
// the model only sees tools whose predicate currently evaluates true
How it works
Plan. Judge. Orchestrate.
Every request runs the same disciplined loop. The harness plans the work, judges multiple attempts, and orchestrates how they run — the model fills in the blanks, the harness owns the process.
Plan
Rehex breaks every request into a visible task DAG. You see the work before it starts.
Judge
Instead of betting one shot is right, Rehex runs several and a candidacy judge picks the winner.
Orchestrate
Parallel sub-agents, cross-domain. Not just coding — Rehex scaffolds any task the way thin harnesses scaffold one.
Why small models
Small, on purpose.
Run it locally
Tiny models run on your own hardware. Higher privacy safeguards, no data leaving the building — and inference costs only the power you draw.
Predictable by design
Repetitive, bounded tasks with outcomes you can expect. Hand them to a small model under the harness and trust the result.
Less reliance, not replacement
Frontier models aren't going away. The goal is to drastically cut the over-reliance on them — reaching for one only when the work truly demands it.
Numbers are coming.
We're putting Rehex-on-open-weights head to head against frontier setups on cost, reliability, and task completion. Real runs, published in full — no cherry-picked screenshots. Join the waitlist to get them first.
Get early access.
Join the waitlist. We'll reach out as soon as early access opens up.
No spam. Unsubscribe anytime.