Context Mode for Claude Code: Do the Cursor and Codex Limits Still Apply?
By Matt Cullerton, founder of Mavric and Apso
Context Mode is a free, source-available add-on that keeps raw tool output out of your coding agent's context window and restores file edits and tasks after a chat compacts. A TikTok from @dailyaisignal says it works fully only in Claude Code, because Cursor drops the context its hook adds and Codex supports deny rules only. The Claude Code part holds up, the Cursor limit was confirmed by Cursor staff through June 2026, and the Codex limit is out of date.
What is Context Mode?
Context Mode is a plugin by Mert Koseoglu, published at github.com/mksglu/context-mode. It hooks into the agent so tool output gets processed before it reaches the model. Per its README, it runs code against tool output in a sandbox so only stdout enters context, and it keeps an FTS5 index with BM25 retrieval over past session events.
On Claude Code it registers six hooks: PreToolUse, PostToolUse, UserPromptSubmit, PreCompact, SessionStart and Stop. All six are in Anthropic's hooks reference. That reference says a PreToolUse permissionDecision can be "allow", "deny", "ask", or "defer", that updatedInput "replaces a tool's arguments before it runs", and that additionalContext "passes a string from your hook into Claude's context window."
Does Context Mode really cut context use by 98%?
Every number in the video comes from the first forty lines of the README: "A Playwright snapshot costs 56 KB. Twenty GitHub issues cost 59 KB." "After 30 minutes, 40% of your context is gone." "315 KB becomes 5.4 KB. 98% reduction."
The repo's BENCHMARK.md gives the fuller picture. It covers 21 scenarios built from "real outputs from popular Claude Code MCP servers, Skills, and dev tools". The 98% figure is the Part 1 subtotal, which the benchmark states as "315 KB raw → 5.5 KB context (98% savings)", so the README's 5.4 KB is a rounding step off. Across all 21 scenarios the result is 376 KB of raw output down to 16.5 KB, which the file states as 96%. The "99% savings" on the video's opening card is a single row, the Playwright snapshot (56.2 KB to 299 B).
The "40% after 30 minutes" figure has no measurement method in the README or the maintainer's February 26, 2026 launch post, so treat it as a claim.
Does Context Mode work with Cursor?
Partly. The README says: "Known limitation: Cursor accepts additional_context in hook responses but does not surface it to the model." It cites Cursor forum thread 155689, opened on 2026-03-23 against Cursor 2.6.20. Cursor's Mohit Jain replied on 2026-03-24 that the field "is accepted and validated by the hook runner" but "the value is never actually injected into the model's conversation context." The last update, on 2026-06-15, reports the same failure on Cursor 3.7.27 and 3.7.36.
Cursor's own hooks documentation says the opposite. It lists additional_context for postToolUse, postToolUseFailure and sessionStart as "Extra context injected into the conversation after the tool result." Cursor staff confirmed the bug, users still reported it in June 2026, and the thread has had no update since.
The README also says Cursor's sessionStart hook is "currently rejected by their validator." In thread 149566, Cursor's Colin answered on 2026-03-30: "This was fixed in v2.4."
Does Context Mode work with Codex?
Yes, on current builds. The video's "deny rules only" line repeats a README note that says Codex routing "currently supports deny rules only" and "still needs upstream updatedInput support," tracked in openai/codex#18491.
That issue is still open, but its author posted on 2026-07-09 that "Both blockers from the original issue appear resolved upstream." The pull request "Support PreToolUse updatedInput rewrites" (#20527) merged on 2026-05-12, and "Default function tools into tool hooks" (#23757) merged on 2026-05-23.
Codex's hooks documentation now says PreToolUse "can intercept Bash, file edits performed through apply_patch, MCP tool calls, and other local function tools." It documents updatedInput rewrites and additionalContext, and says "Hooks are enabled by default." The README still tells Codex users to turn hooks on with a [features].hooks flag.
Context Mode's own Codex hook agrees with the docs. The header of hooks/codex/pretooluse.mjs says it honors permissionDecision:"allow" + updatedInput and additionalContext "on codex-cli >= 0.141.0", with older builds falling back to deny. Codex 0.141.0 shipped on 2026-06-18, four months before the video. Deny-only applies to Codex builds older than that.
Is Context Mode free?
The plugin is free. The license is the Elastic License 2.0, which grants a "royalty-free" right to use, copy, modify and distribute the code but bars offering it "to third parties as a hosted or managed service." The commit history shows a switch from MIT to ELv2 on 2026-03-03, so "source-available" is the right label. The README says "No telemetry, no cloud sync, no usage tracking, no account required."
The same name also covers a paid hosted console. context-mode.com/pricing lists Free, Starter at "$20 /month or $200 a year", Pro at "$49 /month or $490 a year", Team and Enterprise. The README's /context-mode:ctx-insight command opens that console.
How much of this does Claude Code already do?
Some of it. The Claude Code changelog shows version 2.1.261 (2026-09-04) adding bashOutputMaxChars and taskOutputMaxChars, which control how much output Claude receives inline before the rest is saved to a file. Version 2.1.265 added "a 1 GB cap on tool results saved to disk," and 2.1.287 lists "Improved handling of large MCP tool results."
Compaction also keeps more than it used to. Anthropic's context window page says the summary keeps "files examined or modified with important code snippets, errors and how they were fixed, pending tasks, and current work". The sandboxed execution and the BM25 session index are the parts Claude Code does not ship.
One README detail is wrong: it lists "Claude Code v1.0.33+" as the minimum, but hooks arrived in 1.0.38 and the plugin system in 2.0.12.
Bottom line
Context Mode is a real, free plugin, and its headline savings come from the maintainer's own benchmark, where the full result is 96% rather than 98%. The Cursor limit was real through June 2026. The Codex limit applies only to builds older than 0.141.0, according to the plugin's own code.
Sources
- context-mode README - the source of the video's numbers and the Cursor and Codex notes
- context-mode BENCHMARK.md - the 21-scenario benchmark behind the savings figures
- Claude Code hooks reference - hook events and fields
- Claude Code CHANGELOG.md - version history for hooks, plugins and output limits
- Claude Code context window page - what compaction keeps
- Codex hooks documentation - current Codex PreToolUse support
- openai/codex issue 18491 - the tracker the README cites
- Cursor hooks documentation - Cursor's description of
additional_context - Cursor forum thread 155689 - staff confirmation of the Cursor bug
- Cursor forum thread 149566 - the sessionStart fix in v2.4
- Hacker News item 47193064 - the launch discussion behind the README's badge
- context-mode.com pricing - tiers for the hosted console
Published October 18, 2026. Based on a TikTok by @dailyaisignal. Full video notes (markdown)