ContextOS — Context Operating System for AI Coding Agents

A Deterministic Context Operating System via Model Context Protocol (MCP)

Mitigating context window saturation and attention drift in agentic software engineering through syntax-directed AST manipulation, out-of-band execution isolation, in-situ verification, and topological priority scheduling.

Download macOS App · Quick Setup Guide · 中文文档

Background and System Motivation (Context Saturation & Attention Drift)

In complex software repositories, autonomous AI coding agents face a fundamental systems bottleneck: rapid context window saturation and attention drift. Under conventional agent workflows, assistants rely on indiscriminate whole-file dumping, raw build/test stdout feedback, and conversational trial-and-error. This introduces three systemic failures:

1. Attention Budget Dilution & Hallucination: Large volumes of raw build dumps and irrelevant source lines displace critical architectural invariants and type signatures in the prompt;
2. Exponential Cost of Fault Investigation: Truncated test outputs force the agent into recursive log-retrieval turns, compounding cumulative token consumption quadratically;
3. Architectural Blindness & Session Amnesia: Without a verifiable global symbol graph, clearing the context (/clear) or transferring tasks across subagents results in complete state degradation.

ContextOS addresses this as a Context Operating System implemented over the Model Context Protocol (MCP):

Across real-world multi-step benchmarks, ContextOS reduces redundant context consumption by over 90%, stabilizing attention and ensuring long-horizon development convergence.

What changes in daily development?

1. Architecture becomes a Metro Map

Blocks bind to real files, AST symbols, or directory trees (zero ghost blocks allowed). Dependency and resource directories use one bounded tree anchor instead of per-file bookkeeping; Chains represent horizontal subway rails, and typed Links connect transfer stations orthogonally.

2. Intent-Level 2-Turn Loop, Action Slots & In-Situ Diagnostics

• Action Slots & Zero-Alignment Friction: explore automatically provisions actionable target slots [S1], [S2]. The AI doesn't need to guess line numbers or fiddle with verbatim target strings; it simply selects a slot and passes change({ slot: "S1", append: "..." }) or replaces a symbol, completely eliminating parameter alignment failures.
• In-Situ Verification & Instant Diagnostic Frames Extraction:
  • The change step can embed the verification command directly (verify: "npm test"), atomically performing code patching, test execution, and receipt generation in a single turn.
  • Eliminating Extra Log Rounds: When a test fails in conventional workflows, truncated terminal output forces the AI to start 2–3 extra conversation turns just to inspect logs, wasting 20,000–40,000 tokens per incident. ContextOS features a built-in diagnostic extraction engine (extractDiagnosticBlocks) that parses TAP, Mocha, Jest, and SyntaxError frames directly, inlining test names, + actual - expected assertion diffs, and exact file:line:col stacks into the change failure response. The AI sees the exact issue in the same turn with 0 extra log-fetching rounds.
  • Auto-Revert: If tests fail, autoRevert: true instantly restores disk modifications in milliseconds, collapsing the traditional 6–9 round-trip debug cycle into an ultra-fast 2-turn loop (explore ➔ change with in-situ verify ➔ ship).
• Parallel Batching Over Serial Turns:
  • Conventional AI tools encourage conversational, single-file serial editing that balloons context quadratically across turns.
  • ContextOS natively supports multi-concurrency: batch multi-file modifications in edits: [...] arrays and inspect multiple microservices in parallel via inspect({ paths: [...] }), cutting round-trip latency by over 60%.
• Topological Retention Priority:
  • Uses a priority-weighted context budgeting algorithm: next (0) → slots (1) → where (2) → now (3) → slices (4). In large repos, local code previews are clipped first, guaranteeing that the system architecture map and action slots are never truncated.
• Lightweight Real-Time Blackboard & Instant Hydration (.contextos/blackboard.md):
  • The active session state is continuously persisted to a minimal 300-byte (~65 tokens) blackboard. Developers or agents can /clear the chat at any time or hand off tasks across subagents; calling explore resumes full context in ~150 tokens, preventing multi-turn context inflation.

3. Commands run out-of-context

The command gateway behind verify and in-situ change.verify strips ANSI noise, redacts secrets, saves the full sanitized log into .contextos/logs/, and returns a compact receipt with critical diagnostics, reducing terminal noise by over 98%.

4. Code tools operate surgically with deep slicing (inspect)

Multi-language AST engines (compiler-grade parsing for JS/TS/JSX/TSX, Python, Swift, Java, Kotlin, C/C++, C#, Go, Rust, PHP, Ruby) allow VS Code-style symbol search, outline inspection, and surgical reading/editing with automatic symbol re-anchoring. The first-class inspect tool allows on-demand semantic slicing. When answering pure inquiry/comprehension queries ("where is X", "who calls Y"), the OS automatically omits unneeded edit slots and slices, cutting query context by another 50%.

5. Unified Knowledge & Architectural Decisions

Proposals, audits, architectural decisions, and rules live as OS Documents. README.md and README_zh.md appear read-only in the App Knowledge view with images and links intact.

6. Verification with Evidence

Every verify run produces a signed receipt. Checkpoints track pass/fail status per task, giving AI and humans a shared source of truth for what has actually been proven to work.

7. Long-running processes are monitored live

Dev servers, watchers, and background workers are managed by the Process Host and displayed in the Desktop App's bottom-left sidebar with live PID and port tracking.

Quick Start & Setup Options

ContextOS offers two straightforward onboarding routes for effortless setup:

graph TD
    User([Choose Your Setup Route]) --> ChoiceA[Option A: Download Desktop App]
    User --> ChoiceB[Option B: AI Auto-Setup via Prompt]
    
    ChoiceA --> FlowA[Plug & Play · Visual Metro Map · One-Click Editor Injection]
    ChoiceB --> FlowB[Zero Effort · AI Detects Environment & Configures MCP]

Option A: Desktop App (macOS & Windows · Plug & Play · Recommended)

Download the package matching your environment from GitHub Releases:

Setup Steps:

1. Windows: Run the installer .exe or extract the portable .zip and launch ContextOS. macOS: Unzip the downloaded archive and drag ContextOS.app into /Applications.
2. Launch ContextOS, open Settings (gear icon), select your detected AI editor (Cursor / Claude Desktop / Antigravity / OpenCode / Codex), and click Install / Sync Plugin.

[!IMPORTANT] Run ContextOS.app from /Applications, not directly from Downloads, a mounted DMG, or another read-only volume. Plugin synchronization writes user-level files to ~/.contextos and ~/plugins; the app bundle itself is treated as an immutable source.

3. The App injects the ContextOS MCP configuration and unique Skills directly into your editors. Once configured, you can close the desktop App; it does NOT need to stay running.
4. In your AI coding chat, simply activate ContextOS:

   "Write this proposal into ContextOS and start execution" or "Inspect ContextOS and resume development"

[!IMPORTANT] First-time launch on macOS shows "Cannot be opened" or "Unidentified Developer"? ContextOS is an open-source tool without Apple's paid developer certificate notarization. macOS Gatekeeper will block it on first launch by default. You only need to allow it once:

• Method 1 (System Settings · Recommended): Open macOS System Settings ➔ Privacy & Security, scroll down to the "Security" section, and click Open Anyway next to "ContextOS was blocked". Enter your password to confirm.
• Method 2 (Control-Click Shortcut): In Finder, open /Applications, hold Control and click (or right-click) ContextOS.app, select Open from the context menu, and click Open in the confirmation dialog.

Option B: AI Auto-Setup via Prompt (Zero Effort) {#start-in-30-seconds-ai-auto-setup}

If you are already in an AI coding assistant (Cursor / Codex / Claude Code / Windsurf / Antigravity), let the AI configure everything automatically:

Copy and paste this instruction into your AI coding assistant:
"Please read https://github.com/yubinbin32-ops/ContextOS/blob/main/AI_SETUP.md, detect my system environment, and configure ContextOS for me."

What the AI does in the background:

1. System & Client Inspection: If on macOS, asks if you want the native Desktop App (ContextOS.app) and deploys it automatically.
2. Runtime Verification: Checks for Node.js 22+ (or uses the runtime bundled with ContextOS.app).
3. Targeted Precision Injection: Asks which editors you use (Cursor / Codex / Claude Desktop, etc.) and injects only the selected platforms, eliminating duplicate skill noise.
4. Demand-Driven Collaboration Mode: Asks if you need solo local development or team collaboration, configuring local SQLite or Cloudflare D1 accordingly.
5. Project Initialization: Initializes the current project and verifies tools.

Storage Modes: Local & Experimental Cloud Collaboration

ContextOS treats the local workspace project directory as the absolute source of truth (code reads, AST edits, test runs, and logs always run locally):

• Local Storage Mode: Designed for solo development. Architecture data is stored in the project's .contextos/state.sqlite. 100% offline, private, and zero network latency.
• Experimental Cloud Collaboration Mode: Designed for evaluating team collaboration. Connects to a serverless Cloud Hub (Cloudflare D1 edge database) to synchronize architectural topology and progress across teammates and devices. Treat this mode as experimental until its API and operational model stabilize.
  • Multi-Project Strict Isolation: Cloud Hub partitions entities by projectId. A single Cloudflare Worker backs multiple independent repositories cleanly.
• Lossless Two-Way Switching: Switch storage modes anytime by prompting your AI:
  • "Switch current project to cloud collaboration mode" ➔ Local SQLite graph is pushed to Cloud D1.
  • "Switch current project back to offline local mode" ➔ Cloud snapshot is synced back to local SQLite for offline development.

Reproducible ContextOS Benchmark

1. Real Industrial Benchmark: 8 Distributed Microservices

In an industrial stress test across 8 distributed financial clearing microservices (multi-currency double-entry ledger, dynamic FX conversion, sliding-window TTL idempotency, poison transaction dead-letter queue, step timeout and backward compensation Saga, SHA-256 Merkle audit chain, 3-state circuit breaker, and 5-worker concurrent balance contention), the system demonstrated the following performance metrics:

2. 20-Task Comprehensive Dual-Cohort Stress Benchmark

To empirically evaluate system throughput, stability, and token governance under extended workloads, we designed and executed an intensive benchmark spanning 20 comprehensive lifecycle tasks against identical distributed Saga settlement scenarios, comparing a conventional AI assistant (Cohort A Baseline) against ContextOS (Cohort B Intent OS):

• Task Coverage Dimensions: Architecture planning and Phase synchronization, high-concurrency multi-module AST inspection, surgical business logic mutation with in-situ test verification, compile & assertion failure inline diagnostic extraction, external API schema ingestion, zero-loss cross-conversation rehydration via 300-byte blackboard, high-contention race condition validation, rapid code defect localization, multi-file atomic refactoring, speculative optimization with automatic rollback, 8-path concurrent AST outline extraction, out-of-context benchmark execution, and release boundary finalization.

ContextOS Tool Invocation Breakdown (Cohort B):

• change (in-situ verify & atomic rollback): 10 calls (45.5%) — primary workhorse
• explore (intent discovery & slot dispatch): 3 calls (13.6%) — cold start & cross-conversation relay
• inspect (batch AST outlines & signatures): 3 calls (13.6%) — multi-path architectural audit
• verify (sanitized execution & receipt signing): 3 calls (13.6%) — checkpoint integrity proofs
• ops (graph & plan state machine mutations): 2 calls (9.1%) — architectural bindings
• ship (boundary sealing & receipt archival): 1 call (4.5%) — release closure

3. Static Codebase Context Savings Benchmark

The benchmark is generated from the current repository and is intentionally not hard-coded to a historical Block/Chain/Link count. Run it locally to produce the measurements for your checkout:

Run the benchmarks locally:

node scripts/comprehensive-dual-cohort-eval.mjs # 20-task comprehensive dual-cohort stress benchmark
node scripts/benchmark.mjs                      # Static codebase context savings benchmark
node scripts/comprehensive-dev-eval.mjs        # 10-dimension complete capability benchmark

Architecture Evolution

ContextOS has deliberately changed direction as real usage exposed new costs:

• 0.4.x: regex parsing, Ghost Blocks, and a 49-tool MCP surface made architectural memory unreliable.
• V2: introduced real-code Blocks, AST-backed visibility, SQLite plus graph.json, and formal development governance.
• Intent-level architecture: collapsed the public surface to explore / inspect / change / verify / ship / ops, moved orchestration into the OS, and made module metadata derive from the codebase.
• 2.5.0: hardens project identity, release/update paths, graph synchronization, atomic edits, and lifecycle gates.

Read the full decision history →

For Contributors

git clone https://github.com/yubinbin32-ops/ContextOS.git
cd ContextOS
npm ci
npm test
npm run plugin:verify
npm run verify             # full release gate
npm run desktop:build      # macOS + Swift/Xcode

The versioned .contextos/graph.json is the project's portable graph projection.

Contributing · Security · MIT License