INSOMNIQ
Not a tool — a team that learns
LIVE SESSION
$ nightshift run
QUALITY
0%
Scroll

Meet your autonomous team

Each agent has a role, a talent profile, and a track record. They coordinate, delegate, argue, and iterate until the problem is solved — or they call in reinforcements.

NEXUS
Coordinator
Plans every run. Reads past history, picks strategy, adapts mid-flight. The brain that routes everything.
Assigning run #1
ORACLE
Researcher
Queries the Knowledge Base in 3ms. Hybrid BM25 + vector search. Never starts from zero.
140 KB entries indexed
SAGE
Strategist
UCB1-driven pattern selection. Exploits what worked, explores what's untested. Never gets stuck playing safe.
Evaluating 3 strategies
FORGE
Executor
Does the work. Code, research, analysis, reports — anything you can describe as a problem.
Writing output...
LENS
Evaluator
Honest judge. Scores 0–100% against acceptance criteria. Spawns sub-evaluators for complex output.
Quality: 87%
AUDITOR
Watchdog
Watches the run. Catches stuck loops, budget burns, repeated failures. Delivers diagnosis at every replan.
Monitoring 4 signals
INVESTOR
Ambition driver
Holds board meetings when progress stalls. Sets concrete goals. Forces real improvement, not safe loops.
Watching quality curve
LOOP
Iterator
When good isn't good enough. Replans with intent — carries what worked forward, discards what didn't.
Iteration 2/4

Every other tool forgets.

Devin at $9/hour forgets everything. Cursor doesn't remember yesterday. INSOMNIQ runs 50 is fundamentally different from run 1 — because it remembers what worked.

Capability
Devin / Cursor / Copilot
INSOMNIQ
Learns between runs
✗ Stateless
✓ Knowledge Base
Strategy from experience
✗ Same every time
✓ UCB1 + AR patterns
Non-code tasks
✗ Code only
✓ Any problem
Smart retry
✗ Full restart
✓ Blame-driven
User input mid-run
✗ Not supported
✓ Auditor inbox
Self-evaluation
✗ Basic
✓ Sub-evaluators
Exploration drive
✗ Loops on safe path
✓ Investor pressure

9 pillars. One engine.

70 files · 17K lines · 665 tests. Every feature — KB queries, auditing, evaluation, iteration, cost tracking — flows through one unified engine.

01
Knowledge Base
LanceDB + ModernBERT. Hybrid BM25 + vector search in 3ms. Every run feeds it. Every agent reads it before acting.
LANCEDB · MODERNBERT
02
Agent Resources
The HR department. Scores team patterns with UCB1. Mutates agent configs. Learns which model works best per task type.
UCB1 · AR PATTERNS
03
Auditor + Investor
The oversight pair. Auditor watches, diagnoses, holds meetings. Investor sets goals, pushes for ambition when stuck.
OVERSIGHT PAIR
04
Honest Evaluation
Quality 0–100%. Git diff = truth. Spawns sub-evaluators for complex output. Reviews feed back into KB.
QUALITY 0-100%
05
Unified Engine
One function: _solve_pipeline. No separate code paths. Pipeline, iterations, decomposition — same engine.
ONE PATH
06
File-Based Monitoring
Files are the API. status.json, events.jsonl, inbox.jsonl. Any tool integrates. No server needed.
ZERO API
07
Self-Learning
Episodic memory, KB writes, AR updates, Predictor, Librarian — five feedback loops compounding every run.
5 FEEDBACK LOOPS
08
Exploration Drive
UCB1 exploration bonus + Investor pressure. The system takes calculated risks. Avoids safe-play death spirals.
UCB1 EXPLORATION
09
External Integration
Inject hints mid-run. Any backend: Claude, Gemini, Codex, Ollama. AR learns which model wins per task.
MULTI-BACKEND

Growing with the community

This isn't another LLM wrapper. It's an evolving organism — one that grows smarter with every run, every user, every contributed pattern. You can be part of what it becomes.

AGPL-3.0 Python ≥ 3.11 665 tests passing 191 commits Claude · Gemini · Ollama LanceDB + ModernBERT
Coming soon
Investor Center
Cross-project knowledge aggregation. Your Investor learns from every project you've ever run. Public Center lets the community share collective patterns.
Coming soon
Public Knowledge Base
Opt-in shared KB. Strategies, patterns, domain knowledge shared across the community. Your local KB stays private. Public KB grows with every contributor.
Coming soon
Server Mode
Run INSOMNIQ as a service. Submit problems via API, get results back. Deploy on your own infrastructure or use the hosted version.
Start here
What's your hardest problem?
Give it to the swarm. Watch it think, iterate, and learn.
Free, open source, runs on your machine.
17K
Lines of code
665
Tests passing
9
Pillars
Runs to learn from