HiveMind

OS-inspired scheduler for concurrent LLM coding agents.

When you spawn 10 agents, they shouldn't all stampede the API at once. HiveMind sits between the agents and the LLM provider as a transparent HTTP proxy, managing concurrency, rate limits, priority, and resource allocation — the way an OS kernel manages processes competing for CPU.

Quickstart

# Install
pip install hivemind-scheduler

# Start the proxy (auto-detects provider from URL)
hivemind proxy                                          # Anthropic (default)
hivemind proxy --upstream https://api.openai.com        # OpenAI

# In another terminal, point your agents at it
ANTHROPIC_BASE_URL=http://127.0.0.1:8765 claude code   # Claude Code
OPENAI_BASE_URL=http://127.0.0.1:8765/v1 cursor        # Cursor / Copilot / Codex

That's it. Your agents now go through HiveMind. Zero code changes.

The Problem

11 parallel agents, one API key. 3 died from ECONNRESET/502 — classic connection exhaustion. The surviving 8 worked fine. If they'd been staggered by 5 seconds, all 11 would have succeeded.

The problem isn't capacity — it's coordination.

How It Works

Agent → http://localhost:8765/v1/... → HiveMind Proxy → Anthropic / OpenAI / Ollama / Azure
                                            ↑
                                Admission control (condition variable)
                                Rate limit tracking (provider-aware)
                                AIMD backpressure + circuit breaker
                                Token counting (budget enforcement)
                                Provider-specific retry (429/502/529)
                                SSE streaming pass-through

Agents don't know HiveMind exists. They make normal API calls. HiveMind sits in the middle.

Results

Evaluated across 7 scenarios with 5–50 concurrent agents:

Scenario	Without HiveMind	With HiveMind
10 agents, 50 req/min	100% failure	0% failure
11 agents, realistic errors	73% failure	0% failure
20 agents, stress test	100% failure	10% failure
50 agents, extreme	100% failure	0% failure

Install

pip install hivemind-scheduler          # Core
pip install hivemind-scheduler[all]     # + tiktoken + redis

Or from source:

git clone https://github.com/jayluxferro/hivemind.git
cd hivemind
pip install -e ".[dev]"

Usage

Transparent Proxy (recommended)

# Start the proxy — auto-detects provider from URL
hivemind proxy --upstream https://api.anthropic.com
hivemind proxy --upstream https://api.openai.com
hivemind proxy --upstream http://localhost:11434  # Ollama

# Point agents at it
export ANTHROPIC_BASE_URL=http://127.0.0.1:8765
export OPENAI_BASE_URL=http://127.0.0.1:8765/v1

MCP Server

hivemind serve

IDE Integration

Generate config for your IDE/tool:

hivemind setup claude-code
hivemind setup cursor
hivemind setup windsurf
hivemind setup codex
hivemind setup copilot
hivemind setup all         # Show all configs

CLI Reference

hivemind proxy

Flag	Default	Description
--host	127.0.0.1	Bind address
--port	8765	Bind port
--upstream	https://api.anthropic.com	Upstream API URL (provider auto-detected)
--max-concurrency	5	Max concurrent in-flight requests
--min-concurrency	1	Floor for AIMD backpressure
--db	hivemind.db	SQLite database path
--max-retries	3	Max transparent retries on 429/502
--retry-base-delay	1.0	Base retry delay (seconds)
--retry-max-delay	30.0	Max retry delay (seconds)
--latency-target	auto	Latency target in ms for AIMD (auto-detected from provider)
--aimd-increase	auto	AIMD additive increase (auto-detected from provider)
--aimd-decrease	auto	AIMD multiplicative decrease (auto-detected from provider)
--total-budget	unlimited	Global token budget
--agent-budget	unlimited	Default per-agent token budget

hivemind serve

Flag	Default	Description
--upstream	https://api.anthropic.com	Upstream API URL
--max-concurrency	5	Max concurrent requests
--db	hivemind.db	Database path
--total-budget	unlimited	Global token budget
--agent-budget	unlimited	Default per-agent token budget
--max-retries	3	Max transparent retries
--min-concurrency	1	Floor for AIMD backpressure

MCP Tools

Tool	Description
hm.submit	Submit an agent task to the scheduler
hm.batch	Submit multiple tasks at once
hm.status	Check task/queue status
hm.priority	Adjust task priority (low/normal/high/critical)
hm.budget	Set/check token budgets (per-agent and global)
hm.metrics	Scheduler performance stats
hm.config	Tune scheduler parameters at runtime
hm.setup	Generate IDE/tool integration configs

Architecture

Five Scheduling Primitives

#	Primitive	What it does	OS Analogy
1	Admission Control	Concurrency gate — max N requests in-flight	Process scheduler
2	Rate Limit Tracking	Parse x-ratelimit-* headers, pause proactively	I/O scheduling
3	AIMD Backpressure	Latency-based concurrency: low → increase, high → cut	TCP congestion control
4	Token Budgets	Per-agent + global ceilings, warn at 85%, checkpoint at 100%	OOM killer
5	Priority Queue + DAG	Shortest-job-first, dependency tracking, reprioritization	Nice levels + cgroups

Provider Support

Auto-detected from upstream URL:

Provider	Rate Limit Headers	Default Concurrency	Streaming
Anthropic	Yes	5	Yes
OpenAI	Yes	10	Yes
Azure OpenAI	Yes	10	Yes
Google (Gemini)	-	8	Yes
Ollama (local)	-	2 (GPU)	Yes

Optional Features

pip install hivemind-scheduler[tokenizer]     # tiktoken for accurate token counting
pip install hivemind-scheduler[distributed]   # Redis for multi-machine coordination

Evaluation

Run benchmarks against a mock API (no real API credits needed):

python -m evaluation.run_benchmark --quick     # 5 agents, 30 seconds
python -m evaluation.run_benchmark --replay    # 11-agent original scenario
python -m evaluation.run_benchmark --ablation  # Test each primitive individually
python -m evaluation.run_benchmark             # Full suite (all scenarios)

Testing

pip install -e ".[dev]"
python -m pytest tests/ -v

182 tests covering all scheduler primitives (admission control, backpressure with circuit breaker, rate limiting with provider profiles), proxy, streaming, multi-provider integration (Anthropic + OpenAI), tokenizer, distributed backend, and MCP tools.

License

MIT — see LICENSE.