context-mode: 98% context savings and session continuity across 17 AI coding agents

July 29, 2026·AI Newsroom·5 min read·10 sources

context-mode mcp context-engineering context-window ai-coding-agents token-savings+3

Source: https://github.com/mksglu/context-mode · Captured 2026-07-29 via Playwright Chromium · Project license: ELv2.

mksglu/context-mode is a 19,414★ MCP server that combines a 98% tool-output reduction sandbox, SQLite + FTS5 session-continuity across conversation compactions, a “think in code” paradigm, and 17 supported AI coding agent platforms — all under an ELv2 (source-available) license. Per the project’s own BENCHMARK.md: 21 scenarios, 376 KB raw data processed into 16.5 KB of context, for a 96% overall savings with 100% code-example preservation.

What happened

Every MCP tool call dumps raw data into the context window. A Playwright snapshot costs 56 KB. Twenty GitHub issues cost 59 KB. One nginx access log — 45 KB. After 30 minutes, 40% of the context is gone (README, 2026-07-29). mksglu/context-mode attacks this with four integrated subsystems packed into a single npm package registered across 17 AI coding agent platforms.

The repo stands at 19,414★ / 1,400+ forks / 2,108 commits as of 2026-07-29 (live-verified at github.com/mksglu/context-mode). It recently hit the front page of Hacker News. The project lists “Used across teams at” logos for Microsoft, Google, Meta, Amazon, NVIDIA, and Stripe in its README — self-disclosed in the project’s own README, not a customer reference.

The four sub-problems

Per the README, context-mode addresses four concrete pain points of the AI agent context window:

Tool output flood. A Playwright snapshot costs 56 KB, 20 GitHub issues cost 59 KB, a single access log costs 45 KB. After 30 minutes, 40% of the context is gone.
Session-compaction amnesia. When the conversation compacts to free space, the agent forgets which files it was editing, what tasks are in progress, and what the user last asked for.
LLM as data processor. Agents read 50 files into context to count functions when a one-line script that console.logs the result is faster and uses 100× less context.
Prose-style enforcement. Some aggressive-brevity prompts degrade coding and reasoning benchmarks. The project deliberately does not dictate how the model writes its answer.

The four-part solution

Layer	What it does
Context Saving	Sandbox tools (`ctx_execute_file`, `ctx_index`, `ctx_search`) keep raw data out of the context window. README hero: 315 KB raw → 5.4 KB (98% reduction).
Session Continuity	Every file edit, git operation, task, error, and user decision is tracked in SQLite. On compaction, events are indexed into FTS5 and retrieved via BM25 search — the model picks up exactly where it left off.
Think in Code	Instead of 47 `Read()` calls worth 700 KB, the agent writes one `ctx_execute("javascript", ...)` script that `console.log`s the result — 3.6 KB. Mandatory across all 17 clients.
No prose enforcement	The project keeps raw data out of context but never dictates how the model writes its final answer. Brevity, completeness, formatting — the model’s call.

The benchmark numbers

Per the project’s BENCHMARK.md (21 scenarios, 376 KB raw → 16.5 KB context):

Benchmark slice	Scenarios	Raw → Context	Savings
Part 1: `ctx_execute_file` (structured data)	14	315 KB → 5.5 KB	98%
Part 2: `ctx_index` + `ctx_search` (knowledge retrieval)	6	60.3 KB → 11.0 KB	82%
Part 3: Large-output externalization	1	Data >100 KB auto-indexed to FTS5	Pointer only
Full debugging session	—	177.1 KB → 10.2 KB (~45,300 → ~2,600 tokens)	94% more context free

Representative ctx_execute_file rows: Playwright page snapshot 56.2 KB → 299 B (99%), nginx access log 45.1 KB → 155 B (100%), analytics CSV 85.5 KB → 222 B (100%). All 125 tests pass. The benchmark is the project’s own fixture set, not an independent replication.

The 17-platform scope

Full plugin (MCP + hooks + skills + slash commands): Claude Code (marketplace), OpenCode, KiloCode, OMP (Oh My Pi), Antigravity CLI / agy, GitHub Copilot CLI, Codex CLI.

MCP + hooks (manual config): Gemini CLI, VS Code Copilot, JetBrains Copilot, Cursor (work in progress — marketplace awaiting Cursor team review), Kiro, Kimi Code, Qwen Code, OpenClaw / Pi Agent.

MCP-only (no hook support): Zed, Antigravity IDE.

The project counts “17 supported clients, plus the OpenClaw gateway integration.” On Claude Code, the install is two commands: /plugin marketplace add mksglu/context-mode, then /plugin install context-mode@context-mode. All 11 ctx_* tools become available.

The license, in plain language

context-mode is licensed under the Elastic License 2.0 (ELv2) — the source is fully readable, anyone can install it locally, and the project accepts upstream contributions. The hosted-service restriction prevents others from selling context-mode as a managed SaaS. For a developer running npm install -g context-mode and using it as a Claude Code or Cursor plugin on their own machine, ELv2 has no practical impact.

Practical implications for builders

If you hit compaction amnesia on a long Claude Code session, context-mode is the simplest drop-in fix: the SessionStart hook injects the routing block, PreCompact builds a resume snapshot, and the next session’s ctx_search retrieves only the relevant events.
If you burn context on tool output, ctx_execute_file compresses the result to a one-line summary; the full content lives in FTS5, queryable on demand.
If your agent reads 50 files to count functions, write a one-line ctx_execute script instead — 100× less context, same answer.
If you bounce between Claude Code and Cursor, the same context-mode install is reachable from both.
If you need a hosted service, this is not the project for that — ELv2 prevents a managed offering.

Risks and caveats

The 96% / 98% / 82% / 94% numbers are the project’s own BENCHMARK.md, not an independent evaluation.
The “Used across teams at” logos in the README are self-disclosed, not customer announcements.
ELv2 is source-available, not OSI-approved open source. The article says this in plain language above.
context-mode runs user-level MCP tools that read files, run shell commands, and access the network — the same risk profile as any MCP tool.
The Cursor marketplace plugin is work in progress (tracking #485 / #489); the local-folder install path is a documented workaround.
This article covers one tool in a growing context-engineering category. Adjacent tools — Ratel (BM25 tool catalog), Headroom (transparent compression proxy), and codebase-memory-mcp (code-intelligence graph) — solve different sub-problems and are not compared here.

What to watch

The Cursor marketplace plugin (tracking #485 / #489).
Any change in ELv2 terms or a re-license to Apache-2.0 / MIT.
The first independent third-party benchmark on a non-fixture workload.
New client platforms — the list has grown from 2 (Claude Code, Gemini CLI) to 17+.
The next ctx_execute major version — currently on rapid minor-version cadence.