ostk — An Invisible OS for AI Coding Agents
The problem with AI agent frameworks
Most agent frameworks ask you to learn their API, pipe your code through their cloud, and trust their coordination layer. That works until you have two agents editing the same file, a compliance requirement that code stays local, or a budget that can't absorb $200/day in API calls.
What ostk does differently
ostk is an operating system for AI coding agents. Not a framework — infrastructure. A single Rust binary that runsentirely on your machine.
The core insight: agents already know how to read and write files. Instead of teaching them a new coordination API, giving them OS-level primitives that work invisibly underneath.
Five laws define the system:
- The write path is invisible — agents write files normally; the OS intercepts, versions, and audits every mutation
- Agents are ephemeral — state lives in the filesystem, not in memory
- Coordinate through the filesystem — no message buses, no RPC
- Optimistic concurrency — no locks; conflicts resolve at write time through a four-tier "Hot PR" system
- Infrastructure is invisible — if agents notice the OS, it failed
What ships in the box
- Kernel: process supervision, filesystem coordination, conflict resolution, GPG-signed audit trail, output compression, quota enforcement
- Fleet management: spawn agents, monitor heartbeats, reap dead
processes —
ostk spawn,ostk ps,ostk reap - Work orchestration: hay (raw thinking) compiles into needles (executable work with verb + target + test). No Jira required.
- Semantic compression: VTE stripping and deduplication save ~80% of tokens on command output. Agents see signal, not noise.
- Context injection: Silent injection of project context into agent prompts. On needle-bench (1,664 runs across 26 models), this improved debugging performance by +34.3 percentage points on average. Devstral 24B with context hits 75% — outperforming GPT-4.1 bare at 50%.
- TUI console: real-time visibility into agent execution, permission gates, momentum tracking
Who it's for
- Solo developers and small teams running AI agents on their own codebases.
- Security-conscious orgs that can't send code to external APIs. Anyone tired of paying per-token for work a local model can do with the right context.
ostk runs on air-gapped networks. It works with any model that speaks an API — Claude, GPT, Gemini, Llama, Mistral, Devstral, Qwen, DeepSeek.
The evidence
ostk was built by agents running on ostk. Hundreds of needles tracked, compiled, and closed through the OS itself — no separate CI, no project management tool, no Slack. The commit history is the proof.
The needle-bench suite runs 8 real-world debugging scenarios across every model we can get API access to. Results are public, reproducible, and updated weekly.
Get started
curl -fsSL https://ostk.ai/install.sh | sh
cd your-project
ostk init
ostk bootThree commands. No account, no API key to ostk itself, no config file unless you want one. Your model API keys stay in your system keychain — they never enter LLM context.
