fractal

the recursive language model cli agent

A terminal agent that is an RLM. Powered by predict-rlm — our self-harnessed Recursive Language Model runtime.

$ curl -LsSf https://fractal.trampoline.ai/install.sh | sh

fractal · session 8f31c2 ~/deals/acme · depth 0

fractal› go through this 123 page contract, build a complete timeline, and set reminders for the deadlinesRLM turn 1/30 (ok)reasoning: 123 pages won't fit one context — split itpython:class DatedItem(BaseModel): date: datetime.date description: str results = await asyncio.gather(*[ predict( "page: dspy.Image -> items: list[DatedItem]", page=render(page, dpi=80), ) for page in doc])↳ sub-lm 47/123 · page 47 reasoning: read the dates on the page, label what each governs 2 items · 2026-04-01 renewal notice · 2026-06-30 term end ↳ returning items to parentRLM turn 2/30 (ok)reasoning: 31 items collected — sort them, then write the deliverablepython:items = sorted((i for r in results for i in r.items), key=lambda i: i.date)write_file("timeline.md", to_markdown(items))for i in upcoming(items): add_reminder(i.date, i.description)Contract timeline · Acme MSAAcross all 123 pages I found 31 dated items. Here's the chronology, with the upcoming deadlines flagged:2026-01-15 — effective date; initial 12-month term begins2026-04-01 — renewal-notice window opens (60 days)2026-05-31 — last day to give non-renewal noticedeadline2026-06-30 — current term endsFull timeline written to timeline.md; the 4 upcoming deadlines are on your calendar.fractal›

model fable-5 · sub haiku-4.5 verbose off

how it works

The agent is an RLM.

Most agents call a model in a loop. Fractal's loop is the model — predict-rlm recurses, spawning sub-LMs to work the shards of a task that won't fit one context, then folds their results back up. Fractal is a thin UI on top, adding session memory so you can hold a conversation across turns. It's probably the easiest way to get started with an RLM, and to actually understand how one works — by experimenting on your own tasks.

where it shines

Analysis- and context-heavy work.

Reasoning across a big or deep codebase, synthesizing across many files, audits, investigation — anything where the context is the hard part.

Hand off the heavy work

Use it directly in your terminal, or have your main agent hand the heavy lifting to it in headless mode with fractal -p "…". We ship a skill that teaches agents when and how to reach for it.

what you get

Small surface, deep runtime.

◆Powered by predict-rlm

Recursive and self-harnessed. The runtime is the agent — no orchestration to assemble.

◆Model-agnostic

OpenAI, Anthropic, Gemini, Groq, Ollama, OpenRouter, or any OpenAI-compatible endpoint.

◆Sandboxed by default

Every turn runs in an isolated Docker sandbox. Point it at real work without flinching.

◆Headless & scriptable

Drive it from CI or another agent with fractal -p.

go deeper

The real magic is the runtime underneath.

→ predict-rlm

The recursive, self-harnessed RLM runtime that powers Fractal.

→ The RLM paper

Recursive Language Models, from MIT CSAIL.

→ Discord

Build with us. It's early — we'd genuinely love contributions.

Fractal is a fully open-source proof of concept we're putting out to see what people build with it. It's early, and moving fast.