HotCopy.ai

company

https://hotcopy.ai

Activity Feed

AI & ML interests

Recursive Language Models

Recent Activity

hotcopy updated a Space 4 days ago

HotCopyAI/hotcopy-rlm-demo

hotcopy updated a Space 4 days ago

HotCopyAI/README

hotcopy updated a dataset 4 days ago

HotCopyAI/rlm-trajectories-seed

View all activity

Organization Card

Community About org cards

HotCopy is a Recursive Language Model CLI on Cloudflare Workers. Unbounded context, agent swarms in parallel, two-tier orchestrator/worker architecture (256K context per agent), no API keys, no config.

▸ OPS-BRIEF / SECTOR: DEV-INFRA

h.HotCopy.ai

Your codebase is too big for any context window. So we killed the context window. Managed recursive AI coding CLI — agent swarms decompose impossible problems in parallel.

Install

▸ UNBOUNDED CONTEXT // 50+ PARALLEL AGENTS // 0 API KEYS

What HotCopy is

Managed recursive AI coding CLI. Agent swarms decompose impossible problems in parallel. No context limits. No API keys. No config. Just install and go. HotCopy treats your project as an environment the model interacts with programmatically — never raw files in a prompt.

The Recursive Language Model paradigm

HotCopy implements the Recursive Language Model (RLM) paradigm introduced by Zhang, Kraska, and Khattab at MIT CSAIL (arXiv:2512.24601, Jan 2026). Existing scaffolds (Claude Code, Gemini CLI, Cursor) place the user prompt directly into the LLM's context window, generate output autoregressively (capped by token limits), and verbalize sub-LLM calls (only O(1) delegations). RLMs invert this: the prompt is loaded as a REPL variable, the model writes code to interrogate it, sub-calls are programmatic functions inside that code, and the answer is built up in variables and returned via FINAL_VAR() — unbounded by any single model's generation length.

Benchmarks vs base model

Benchmark	Input Size	Base GPT-5	RLM(GPT-5)	Avg Cost
BrowseComp+	6–11M tokens	0%	91.3%	$0.99
OOLONG	131K tokens	44%	56.5%	$0.43
OOLONG-Pairs	32K tokens	0.04%	58.0%	$0.33
CodeQA	23K–4.2M tokens	24%	62%	$0.11

Median RLM cost ≤ base model cost; up to 3× cheaper than summary-agent baselines. Source: Zhang, Kraska, Khattab. Recursive Language Models. MIT CSAIL. arXiv:2512.24601, Jan 2026.

How it works

Prompt as environment. Project context is loaded into Durable Object memory as REPL variables. The orchestrator sees only metadata (file count, repo map, total chars) — never raw files.
Two-tier orchestrator/worker architecture. The RLM root model owns a 256K orchestration window. Every sub-call routes to a separate worker model with its own 256K window and multimodal capability — managed for you, no provider keys to wire up.
REPL with stdout truncation. The orchestrator writes JavaScript that runs in a sandboxed V8 isolate (@cloudflare/codemode). Stdout is truncated before being appended to history — variables hold the long state, the model history holds the metadata.
Parallel sub-calls via llm_batch(). Unlike the paper's sequential implementation, HotCopy fans out scouts in parallel through Cloudflare Workers, dramatically reducing latency on map-reduce decompositions.
Durable Objects + AI Gateway. Every model call routes through hotcopy-gateway for cost tracking, prefix caching (per-session affinity), and rate-limit handling. State lives in DO SQL — trajectories, sub-calls, costs, all logged for debugging.
FINAL_VAR() for unbounded output. The model returns the name of the variable holding the answer. The reply is whatever that variable holds — no autoregressive output cap.

Architecture deep-dive

REPL scope, three-tier memory, and the permissions model

REPL scope (root code)

Category	Functions
Context	`context.read(path)`, `context.readAll(paths)`, `context.search(pattern)`, `context.preview(chars)`, `context.metadata`, `context.files`
Sub-calls	`llm_query(prompt)`, `llm_batch(prompts)`
Agents	`spawn_worker(role, task, ctx)`, `vision_analyze(base64, prompt)`
Memory	`memory_search(query, limit)`, `memory_failures(query)`, `memory_save(type, content)`
File tools	`write_file(path, content)`, `edit_file(path, old, new)`, `shell(cmd, cwd)`
Verification	`verify(claim, evidence)`, `verify_claims(...)`
Persistent state	`set_context(label, content)`, `load_context(label)`
Terminal	`FINAL(text)`, `FINAL_VAR(varName)`, `print(...)`

Three-tier memory

Working memory — per-agent context window. Compaction at 70–75% utilization. Budget: 12% system, 18% tools, 25% retrieved context, 25% history, 5% user input, 15% output reserve.
Session memory — SWARM_STATE.md task manifest, continuously updated by the orchestrator. Exploits transformer recency bias by placing the live task list at the end of context.
Project memory — D1 + Vectorize. Observations, decisions, failures indexed for semantic search. AI-compressed after N turns. Exposed via MCP for external tool access.

Permissions

Default-deny on fetch_*, browse, crawl, shell. Wildcard rules in D1 (permission_rules) let users selectively allow patterns (api.github.com/*, git status, etc.) at three persistence tiers — session, user, or global. Every quarantined call hits a tier-cascade evaluator; first match wins; "ask" rules round-trip to the CLI for a per-call decision.

Stack

Cloudflare Workers + Agents SDK · Durable Objects (RLMOrchestrator, WorkerAgent, VisionAgent, MemoryAgent) · D1 · Vectorize · R2 · Dynamic Workers (@cloudflare/codemode) · AI Gateway (hotcopy-gateway).