GitHub - mksglu/context-mode: Free, local plugin that keeps tool output out of your coding agent's context window. Sandbox and SQLite FTS5 search, 99% fewer bytes in context. Works with Claude Code, Codex, Pi, OpenCode, Gemini CLI, Copilot CLI and 4 more agents. Zero telemetry.

A free, local plugin for 10 coding agents. It keeps tool output out of your agent's context window and keeps all usage data on your machine. The plugin is free forever.

33K GitHub stars · 683K npm downloads (read 2026-11-01, stats.json) · on the Hacker News front page and GitHub Trending: charts and sources

The Context Mode plugin is free and local, with its source open on GitHub. Context Mode Gateway is a separate hosted service with its own account and subscription. It takes the same idea from tool output to the whole conversation: Gateway Recall keeps long sessions searchable, and Gateway Memory shares what your agents learn across sessions and tools. See how the Gateway works

Contents: Why · Adoption · Install · Try it · Use cases · Benchmarks · How it works · Tools · Security and privacy · Context Mode Gateway · License · Community

Why

A coding agent reads files, runs commands and fetches pages, and every result goes into its context window. The Context Mode plugin runs that work in a local sandbox, keeps the full output in a SQLite index on your machine, and gives the agent the part it asks for.

• The plugin is free forever. Use it at home and at work, free of charge, under a source-available license.
• Runs on your machine. The hooks, the sandbox and search run as local processes next to your agent.
• The plugin keeps your data on your machine. Every index lives in ~/.context-mode: one context.db per project.
• Zero telemetry. The plugin keeps all usage data on your machine. Security and privacy lists each network call it makes.
• Works offline after install. Sandbox runs, file indexing and search use local SQLite and your own runtimes.
• Thinking in Code. The agent writes a short script instead of reading the data, and only what the script prints reaches the context: see How it works, step 5.
• Small replies. On the 36 bench rows with 10 KB or more of raw output, the plugin's median reply is 99.4% smaller than the raw output.

Adoption: stars and npm downloads

Open the interactive chart and point at a bar to read its value. Tabs there show days, weeks, months and years. Stars come from the GitHub GraphQL API (starredAt); npm downloads come from api.npmjs.org/downloads/range. The total up to 2026-10-29 is 683K npm downloads (stats.json keeps a daily copy). A daily job redraws both charts.

• Hacker News. "MCP server that reduces Claude Code context consumption by 98%" reached the front page with 570+ points and 107 comments on 2026-02-28 (thread).
• GitHub Trending. A bot that watches GitHub Trending logged the repo there on 14 days from 2026-04-24 to 2026-10-17 (uibm/trending-watcher).

Starred by engineers at these companies: each name is the company field of 2 or more stargazer profiles, read with the GitHub GraphQL API on 2026-11-01 (data).

Install

Pick your agent: Claude Code · Codex · Gemini CLI · Qwen Code · Pi · OpenCode · Copilot CLI · Kiro · VS Code · ZCode

Open its block and copy its own install command. Each block also holds the update, check and remove steps, and the Install guide walks through them.

The plugin needs Node 24.15.0 or later, from the engines field in package.json. On Windows, add Git for Windows: the shell tools run in Git Bash, and doctor checks for it. contextmode installs as the npm package and command context-mode. The npm commands use @next, the current 2.0 release tag; the plain package name gives the 1.x line. Claude Code, Codex, Gemini CLI and Qwen Code install from GitHub, and each of their commands names the release branch.

claude plugin marketplace add mksglu/context-mode#release
claude plugin install context-mode@context-mode

# update
claude plugin update context-mode@context-mode
# check
claude plugin list
# remove
claude plugin marketplace remove context-mode

codex plugin marketplace add mksglu/context-mode@release
codex plugin add context-mode@context-mode
# one-time step: trust the plugin hooks
npm i -g context-mode@next
context-mode setup --agents codex

# update
codex plugin marketplace upgrade context-mode
# check
codex mcp list
# remove
codex plugin remove context-mode@context-mode
codex plugin marketplace remove context-mode

gemini extensions install https://github.com/mksglu/context-mode --ref release

# update
gemini extensions update context-mode
# check
GEMINI_CLI_TRUST_WORKSPACE=true gemini mcp list
# remove
gemini extensions uninstall context-mode

qwen extensions install https://github.com/mksglu/context-mode:context-mode --ref release

# update
qwen extensions update context-mode
# check
qwen extensions list
# remove
qwen extensions uninstall context-mode

pi install npm:context-mode@next

# update
pi update --extensions
# check
pi list
# remove
pi remove npm:context-mode@next

opencode plugin add context-mode@next

# update
opencode plugin update
# check
opencode mcp list --pure          # OpenCode 1.19
opencode mcp list --standalone    # OpenCode 2
# remove
opencode plugin remove context-mode@next

npm i -g context-mode@next
context-mode setup --agents copilot-cli

# update
context-mode upgrade
# check
context-mode doctor --agents copilot-cli
copilot mcp list
# remove
context-mode setup --remove --agents copilot-cli

npm i -g context-mode@next
context-mode setup --agents kiro

# update
context-mode upgrade
# check
context-mode doctor --agents kiro
# remove
context-mode setup --remove --agents kiro

npm i -g context-mode@next
context-mode setup --agents vscode

# update
context-mode upgrade
# check
context-mode doctor --agents vscode
# remove
context-mode setup --remove --agents vscode

npm i -g context-mode@next
context-mode setup --agents zcode

# update
context-mode upgrade
# check
context-mode doctor --agents zcode
# remove
context-mode setup --remove --agents zcode

Moving Claude Code, Codex, Gemini CLI or Qwen Code from the main branch to the release ref? Follow Already installed from GitHub?.

The npm route works for every agent:

npm i -g context-mode@next
context-mode setup

Setup is also the repair. It shows a plan, saves your earlier config, then applies the plan; when a step ends early, it restores the saved config. To uninstall:

context-mode setup --remove
npm rm -g context-mode

Restart the agent, then run /ctx-doctor in it. Each check prints PASS or a fix command, and the Doctor guide explains each check.

Your first task with the plugin, step by step: First session. Or paste a prompt from Try it right now.

Try it

Paste one of these prompts into your agent after the install. Each one ran once, live, in Claude Code 2.1.312, and its answer matched the source data (evidence and rerun steps). The headline numbers in Benchmarks use the median of 5 runs. Other agents do the same job with the same tools: ctx_execute, ctx_index and ctx_search. After a prompt, run the stats skill and read the report yourself (See the saving).

On https://react.dev/reference/react/useEffect, show me the example that connects to a chat server and cleans up, and list the rules for the dependency array.

The hook sends WebFetch to the sandbox, and the agent searches the page for the example and the rules. 601,984 B raw, 20,860 B reached the context, 96.5% saved (2026-11-02, 2.0.0-next.26, claude-haiku-5-5).

Fetch https://jsonplaceholder.typicode.com/photos. Which photo has the longest title (give its id and title), and how many photos have a title that contains the word "est"?

The agent runs code over 5,000 JSON records in the sandbox and gets back two answers. 1,039,898 B raw, 393 B reached the context, 99.96% saved (2026-11-02, 2.0.0-next.26, claude-haiku-5-5). The answer is two short facts, so this is the best case of the four.

Here is a Linux syslog sample: https://raw.githubusercontent.com/logpai/loghub/master/Linux/Linux_2k.log. How many authentication failures does it contain, and which 3 remote hosts caused the most?

The agent counts failures in a 1,999-line syslog in the sandbox. 216,485 B raw, 566 B reached the context, 99.7% saved (2026-11-02, 2.0.0-next.26, claude-haiku-5-5).

What are the top 10 stories on the Hacker News front page right now? Give title, points and domain for each.

The hook sends WebFetch to the sandbox, and the agent reads ten stories from the page. 35,087 B raw, 1,501 B reached the context, 95.7% saved (2026-11-02, 2.0.0-next.26, claude-haiku-5-5). The front page changes, so the raw size changes with each run.

See the saving

Run the stats skill right after a prompt. Each agent uses its own command; the skill calls ctx_stats and shows its report unchanged.

The first lines of a real report, after the JSON prompt above ran as one ctx_execute call:

context-mode stats
────────────────────────────────────────────────────────────────────────────────

This session · 1 call · <1m active
  1015.5 KB became 3.7 KB: 1011.8 KB kept out (99.6%)
  Without context-mode |████████████████████████████████████████|  1015.5 KB
  With context-mode    |█░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░|     3.7 KB
  about 259K tokens kept out, 1.3 times a 200K window

Captured 2026-11-03 with 2.0.0-next.35: release-out/server.mjs over MCP in a temp HOME.

Next, the same numbers for 20 real tools, in Use cases.

Use cases

Real tool output from public projects, run through Context Mode. Bench of 2026-11-01 on 2.0.0-next.23 (commit 16a4bcc5), median of 5 runs.

Source: bench/results/2026-11-01-darwin-arm64.json, one row per line; SOURCE.md names where each input comes from. Click a task to open its case: the command, the input file with its sha256, and the run command, so you can run the same row on your machine.

The agent gets the part it asks for. The full output stays in the local index, and ctx_search reads it back. Rows with a command run in the sandbox; the agent then finds the rest with search. The Context7 row has 6,917 B of raw output, under the 10 KB line of the headline rule, so the headline leaves it out.

Next, how the bench counts these bytes, in Benchmarks.

Benchmarks

Rule for every headline number in this README: rows with 10 KB or more of raw output, median of 5 runs per row, measured with ctx_bytes against raw_bytes; a category is the median of its rows; 1 KB = 1,024 B. All 36 such rows, the six categories below: the plugin keeps a median of 99.4% of output bytes out of the context. Bench on 2.0.0-next.23 (commit 16a4bcc5 and the local edits that the file records), macOS arm64, 2026-11-01.

Search replies (category b) carry the matching text, so they save 77.2%. Category e (long sessions) is in Longer sessions. The full file lists every row, small ones too: 2026-11-01-darwin-arm64.md. Reproduce with cd bench && node run.mjs. It needs plain Node and runs offline; bench/README.md lists each scenario and its options. Method, data files and rerun steps: Benchmarks guide.

Longer sessions

Say you debug a slow test all afternoon. The agent reads files, runs the suite and fetches docs, and each result goes into the context window. When the window is full, the host compacts it, and from then on the agent sees a short summary plus your next message.

Context Mode keeps the full output of each large tool call in context.db on your machine and gives the agent a small reply. At compaction, a snapshot of at most 2,048 B keeps the session state. After the compaction, the agent calls ctx_resume to read that snapshot back and ctx_search to find any earlier output by keyword.

In the bench, a scripted session of 40 tool outputs and one compaction made a 442,893 B transcript, and the ctx_resume block that brought the agent back was 2,019 B (row e-long-session in the bench file; open the case). In the same bench, one large log of 666,796 B put 3,380 B into the context, and the full log stayed in the index for ctx_search (row d-synth-1m; open the case). How the snapshot picks what to keep: the Long sessions guide and ADR-0009.

Speed

Median of 5 runs, bench of 2026-11-01.

Source: rows g-hook-kiro to g-hook-claude-code, g-search-latency and g-index-per-mb of the bench file, with every agent and query. The overhead cases show how each one runs.

Next, what happens inside one tool call, in How it works.

How it works

One tool call, in time order, from the install to a hook error.

1. Install. Each agent's own install command adds the MCP server, the hooks and the skills ctx-doctor, ctx-stats and ctx-upgrade to that agent's config folder, for example ~/.codex/skills/<name>/SKILL.md. context-mode setup writes or repairs any part the native command leaves out (ADR-0023).

2. Session start. The SessionStart hook opens the project database. When the session ran before, the agent gets one line that tells it to call ctx_resume (dispatch.ts).

3. PreToolUse checks. Before a tool runs, the PreToolUse hook checks the call in a fixed order: the user's deny list, WebFetch, curl and wget, inline HTTP and build commands, then a full-file Read over OUTPUT_CAP (16,384 B) and one-file shell dumps such as cat or tail. A Read with a line range runs as usual (decide.ts).

4. The exact call. When a check matches, the hook stops the call and its reply names the exact ctx_* call to use, for example ctx_execute with {"language":"shell","code":"<the command>","intent":"<what you need>"}. For WebFetch the reply names ctx_index, then ctx_search (read-deny.ts).

5. The sandbox. This is the Thinking in Code step: the agent's script runs here, and only what it prints comes back. ctx_execute writes the script to a temp folder and spawns the runtime from an argument list in the project folder. The timeout is EXEC_TIMEOUT_MS (600,000 ms, default and maximum); on timeout it stops the whole process group (taskkill /F /T on Windows). The child gets your environment minus preload variables such as NODE_OPTIONS, and the agent gets what the script prints, at most OUTPUT_CAP (execute.ts). Thinking in Code. The agent sends a short script as a tool parameter, the script reads the data in the sandbox, and only what it prints goes into the context. Two real calls, in the form the Claude Code terminal shows them:

   ctx_execute on 5,000 JSON records, from the Claude Code 2.1.312 stream of the Try-it run with Context Mode 2.0.0-next.26 (04-api-json.json). The data is 1,039,898 B; the agent gets 254 B.

   ⏺ context-mode - ctx_execute (MCP)(language: "javascript", c…jsonplaceholder.typicode.com/photos\");
     if (!res.ok) throw new Error(\"HTTP \" + res.status);
     const photos = await res.json();
     let longest = null;
     for (const p of photos) {
       if (!longest || p.title.length > longest.title.length) longest = p;
     }
     const withEst = photos.filter(p => p.title.includes(\"est\")).length;
     const ties = photos.filter(p => p.title.length === longest.title.length).map(p => p.id);
     console.log(JSON.stringify({ total: photos.length, longest…nd count of titles containing \"est\"")
     ⎿  exit 0 · 239 B
        {
          "total": 5000,
          "longest": {
            "id": 653,
            "title": "temporibus commodi et fugit doloremque re…rit exercitationem necessitatibus aut",
            "length": 86
          },
          "longestTieIds": [
            653
          ],
          "titlesContainingEst": 892
        }
   

   ctx_batch on an install log and a long git log, one live MCP call to release-out/server.mjs of Context Mode 2.0.0-next.36 on 2026-11-03 (2026-11-03-ctx-batch.json). The commands print 228,881 B; the agent gets 8,667 B, the parts that match its two queries, and the full output stays in the index.

   ⏺ context-mode - ctx_batch (MCP)(commands: ["cat npm-install… queries: ["cache miss","feat commit"])
     ⎿  ## cache miss
        js.org/@esbuild%2fwin32-x64 282ms (cache miss)
        npm http fetch GET 200 https://registry.npmjs.org/@esbuild%2flinux-s390x 79ms (cache miss)
        npm http fetch GET 200 https://registry.npmjs.org/@esbuild%2fopenbsd-x64 101ms (cache miss)
        npm http fetch GET 200 https://registry.npmjs.org/@esbuild%2fwin32-arm64 99ms (cache miss)
        npm http fetch GET 200 https://registry.npmjs.org/@esbuild%2fdarwin-arm64 99ms (cache miss)
        … +95 lines (ctrl+o to expand)
   

6. Indexing. The full output goes into the index in chunks of at most MAX_CHUNK_BYTES (4,096 B). A Markdown chunk stays inside one heading section, and a code block that fits stays whole. Each chunk goes into two SQLite FTS5 tables: a Porter-stemmed table for words and a trigram table for paths, names and parts of words (fts.ts).

   ctx_index on a copy of this repo's docs/how-it-works.md, one live MCP call to release-out/server.mjs of Context Mode 2.0.0-next.36 on 2026-11-03 (calls.json). The file is 11,569 B; the agent gets 64 B.

   ⏺ context-mode - ctx_index (MCP)(path: "how-it-works.md")
     ⎿  Indexed 1 source(s), 17 chunk(s) from s1. Label: how-it-works.md
   

7. Search. ctx_search ranks each table with BM25 and joins the two lists with Reciprocal Rank Fusion: a hit scores 1/(RRF_K + rank), with RRF_K = 60. Before it returns an empty list, it tries a spelling fix from the indexed words. Search covers this session, this worktree (the default) or the whole project, and each snippet is at most SNIPPET_MAX_BYTES (800 B) (search.ts).

   ctx_search on that index, in the same run, in a project that also holds an AGENTS.md. One search reads the index and every agent's memory files. The agent gets 1,268 B: one content hit and one memory hit, each with its source and heading. For memory shared across machines, agents and sessions, see Context Mode Gateway Memory.

   ⏺ context-mode - ctx_search (MCP)(queries: ["compaction state block"])
     ⎿  Query "compaction state block":
        how-it-works.md #14: ### Session continuity
        
        The hooks log each prompt, edit, command, tool error, i…e same rows always give the same bytes.
        memory AGENTS.md #0: # Agent rules
        … +10 lines (ctrl+o to expand)
   

8. Run, then replace. On hosts that run the tool first, the PostToolUse hook checks the result. Output over REPLACE_TRIGGER_BYTES (4,096 B) goes into the index whole, and the agent gets the best-matching chunks and one ctx_search line for the rest (dispatch.ts).

9. Journal and compaction. The hooks log each prompt, edit, command, tool error and decision in a journal. After a compaction the agent gets a state block of at most STATE_BUDGET (2,048 B), and each long line becomes a ctx_resume(seq) pointer to the full row; one ctx_resume reply is at most RESUME_CAP (2,048 B). On the next prompt, a prompt-match block brings back the journal rows that match it by BM25 (ADR-0009).

   ctx_resume with empty parameters returns the state block, in the same run: 209 B.

   ⏺ context-mode - ctx_resume (MCP)()
     ⎿  <context_mode_state session="mcp:85e67fa2-28ef-45eb-9d5b-a967129b74c0" upto="2">
        indexed how-it-works.md
        source s1: how-it-works.md (11536 B)
        more: ctx_search(queries:[], scope:"session")
        </context_mode_state>
   

10. The database. Each project has one context.db under ~/.context-mode, in Node's built-in node:sqlite. The database runs in WAL mode and runs a PASSIVE checkpoint on close. A new database gets built under a temp name and moves into place in one atomic step, and the busy timeout is DB_TIMEOUT_MS (5,000 ms) (db.ts).

11. A hook error. A hook always ends with exit 0. On any error it gives the host's allow reply and writes one log line to ~/.context-mode/hook-errors.log, so your agent keeps working. A spawned hook has HOOK_TIMEOUT_MS (5,000 ms) (fallback.ts).

Every number with its reason: How it works. The rules the hook follows: ADR-0008. Next, the Tools the agent calls.

Tools

Three more real calls, from the same live run of Context Mode 2.0.0-next.36 on 2026-11-03 (calls.json).

ctx_stats: the agent gets 3,097 B, the first line is the summary.

⏺ context-mode - ctx_stats (MCP)()
  ⎿  ctx_stats: 82.5% kept out all-time, 82.5% this session.
     … +59 lines (ctrl+o to expand)

ctx_doctor after context-mode setup --agents claude-code: 862 B, one row per check.

⏺ context-mode - ctx_doctor (MCP)()
  ⎿  Environment: context-mode 2.0.0-next.36 (8facf27), agent claude-code 0 (client r20-live)
     Environment: darwin-arm64, Node v26.8.1, shell /bin/zsh
     | Check | Status | Detail | Fix |
     |---|---|---|---|
     | D1 version | PASS | 2.0.0-next.36 (8facf27): ahead of next on npm (2.0.0-next.34) |  |
     | D2 node | PASS | ok |  |
     … +11 lines (ctrl+o to expand)

ctx_purge deletes one source: 66 B.

⏺ context-mode - ctx_purge (MCP)(source: "how-it-works.md")
  ⎿  ctx_purge: deleted 18 rows (scope: source). This cannot be undone.

Full reference: Tools.

Commands in the agent: /ctx-stats shows the ctx_stats result, /ctx-doctor runs ctx_doctor (send its report with a defect report), and /ctx-upgrade runs ctx_upgrade, then ctx_doctor.

Security and privacy

• The plugin keeps all its data on your machine, one database per project in ~/.context-mode, and ctx_purge deletes it by session, source or project. Read the Security guide.
• The plugin has zero telemetry: usage data, events and crash reports stay on your machine. It uses the network for five calls, each one started by you or your agent:

• The guard test D-PRIVACY-NET-1 in tests/0008-enforcement.test.mjs scans the source for HTTP clients, sockets, DNS lookups, git network calls and gh or npm spawns. It flags each new call site, so this table stays true.
• Doctor and the debug report remove secrets by their structure before they print.

Code from ctx_execute runs on your host as your user, with your files, your network and your environment except the preload variables, so keep your agent's own permission prompts on. Here "sandbox" means the output stays in the sandbox and out of the context. What the plugin stores, where, and how to delete it: Privacy.

Context Mode Gateway

The plugin works on tool output, on your machine, and it is complete on its own. Context Mode Gateway is a separate hosted service with its own account and subscription. It works on every request, across all your agents. Its 11 features, the full list: Context Saving, Recall, Memory, Skills, Search, Session resume, Active Context, Instructions, Protect, Prompt caching, Live and Analytics. Add the Gateway when you want savings on every request, memory across sessions and agents, or cost reports for a team. For one project on one machine, the plugin covers the job. The two work together: Gateway guide. Gateway Recall, Search and Session resume run on the Gateway server, across sessions and agents. The plugin's ctx_search and ctx_resume read the local context.db of one project. With the Gateway on, your requests pass through its server; the plugin keeps its own data on your machine.

Set up the Gateway with one command:

npx @context-mode/cli signs you in to your Gateway account, which has its own subscription, and writes one config change per agent. Apps: Claude Code, Claude Desktop, Codex, Codex Desktop, T3 Code, OpenCode and Pi. The Gateway CLI needs Node 20.12 or later, from its npm engines field.

Live. One session as it runs: money saved this session, bytes kept out per request, history kept and sent, and a feed of each request with its cost and saving. Click the image for the full video.

Docs: How saving works

Numbers in these screens come from the author's own account over 30 days (Oct 3 - Nov 1, 2026), so read them as one account's usage. Click an image for the full size.

Three prompts to try with the Gateway on. The line under each one names the feature it shows.

Move this project from Jest to Vitest. Convert every test file, fix the config, run the suite after each batch, and tell me what is left.

Memory keeps the plan and the file list across a long job.

Pick up where we stopped yesterday on the Jest to Vitest move. What is done, what is left, and which test was failing?

Recall brings back yesterday's work in a fresh session.

Which date library did we choose on Tuesday, and why did we reject the other two?

Search finds the answer in past sessions.

Gateway only: its cost against Claude Code alone, Claude Opus 5.5, mean dollars per session at list price. The plugin numbers are in Benchmarks.

• Six short tasks (2026-09-29): 24.5% fewer dollars pooled over 176 sessions (21% to 29% per task). v3-opus55-short.json
• Long Claude Code session, 200K window (2026-10-03/04): mean cost per session $4.8627 to $3.0353 (-37.6% in dollars, mean of 3 pairs). ab.json
• We fixed the rules before each run: contextmode.ai/docs/benchmarks.

The Gateway routes to Anthropic and OpenAI. The Gateway sends your requests to its server on the way to the model. It keeps a searchable record of your sessions for Recall and Search, and masks secrets and e-mail addresses when it reads archived output back. Its privacy page is at contextmode.ai/privacy. Keep your agent's own sandbox on next to Protect.

License

Elastic License 2.0 (ELv2, the LICENSE file), a source-available license. You may use, copy, change and share the code, and use it at work, free of charge. The license keeps three uses for the licensor:

1. Offering the software to others as a hosted or managed service that gives them a large part of its features.
2. Moving, changing, turning off or getting around the license key functions, or removing or hiding what the license key protects.
3. Changing, removing or hiding the licensor's license, copyright or other notices.

We chose ELv2 so the source stays open to all and the hosted service stays with the project.

Community

Report a defect on Discord with the full /ctx-doctor report, and send ideas there too. The maintainer writes every change; CONTRIBUTING.md shows how an agent files an issue.

Videos. Better Stack on Context Mode. Make your own video and share it on Discord; we add videos over 10K views here.