What each product actually is
Caveman started as a skill that makes agents answer in terse caveman speech. Version 4.0 of the CLI (4.0.1 on npm, October 13, 2026) bundles six parts that are all on after install: Say less (short answers, code kept exact), Read less (logs, JSON, CSV, YAML and test output compressed before the model reads them), Auto (model and effort routing per ask), Blocks (the agent keeps and reruns scripts it writes), Waste fixes and Browse (compressed web pages). A separate SDK and middleware bring the compression into agents you build, with 21 adapters across TypeScript and Python. Caveman Cloud is the paid layer for observability, evals, agent runs and routing.
Context Mode describes itself as three products. The main one is a hosted gateway: Claude Code or Codex gets its base URL pointed at gateway.context-mode.com, and the gateway folds tool output, trims tool schemas, adds memory and skills, checks Protect rules and records tokens and cost before calling Anthropic or OpenAI with your own credential. Context Mode Engine is the free context-mode plugin and MCP server where the project began; it keeps raw tool output in a local sandbox and index. Context Mode Insight reads Engine events for team analytics.
So the closest like-for-like pairs are Caveman's local CLI against Context Mode Engine, and Caveman plus Caveman Cloud against the Context Mode gateway. The two vendors overlap on token reduction but differ on where the work happens.
How each one cuts tokens
Caveman works on both sides of the conversation. Say less shortens the agent's prose while leaving code, commands and errors byte for byte. Read less shrinks eligible tool output and keeps the original on disk so the agent can ask for it again. Auto lowers cost by choosing a cheaper model or lower effort for easy asks. Blocks targets the tokens an agent spends retyping the same scripts.
Context Mode takes the opposite view on output. The Engine README says it keeps raw data out of context but never dictates the final answer's style, and it argues that aggressive brevity prompts can hurt coding benchmarks. The Engine runs tool work in sandboxed subprocesses in 12 language runtimes, so only printed output enters the context, and indexes larger output in SQLite FTS5 with BM25 search. It also tracks edits, tasks and decisions so a session can continue after compaction. The gateway applies the same idea on the request path: it folds test, build and log output the first time the model sees it, archives the full output, and keeps folded results stable so the prompt cache keeps reading them.
Where your requests and code go
This is the clearest architectural difference. Caveman's README says the skill runs on your machine and sends nothing, except that /caveman-compress sends the file you compress to your configured model provider. Requests from the CLI go from your machine to your provider on your own key. Auto is the exception: once you sign in and pick it, each ask sends Caveman Cloud your latest ask with whatever the agent attaches, the previous ask, the end of the last reply and some request metadata, and on the Free plan Caveman may keep that to improve routing.
The Caveman CLI also sends usage statistics by default: commands run, token counts, a random install ID, OS and IP, but not prompts, code or file paths, according to the README. It can be turned off with caveman telemetry off or DO_NOT_TRACK=1. On Caveman Cloud's Free plan, the pricing page says Caveman may use prompt text and recorded requests to improve its models; the Team plan turns content sharing off, and Enterprise runs in your own VPC.
Context Mode's gateway is on the request path by design: every request passes through its Cloudflare-hosted service, which records tokens, cost and savings and keeps an archive of tool output and sessions for Recall and Search. Its docs say secrets and e-mail addresses are masked before storage, and Enterprise can run in a dedicated instance or your own Cloudflare account. The Engine is the opposite: its README states that nothing leaves your machine, with no telemetry, no cloud sync and no account, unless you connect Insight.
Supported agents and setup
Caveman installs with one npx command on macOS, Linux and Windows and needs Node.js 22.13 or later. It wraps Claude Code, Codex, Gemini CLI, OpenCode and, per its README, 30 or more agents in total; support for short replies and for context compression differs by agent. Auto works in Claude Code, OpenCode and Codex.
The Context Mode gateway currently supports Claude Code and Codex only; OpenCode and Pi are marked as coming soon. Setup is npx @context-mode/cli, which signs you in, edits one config file per agent and keeps a backup; removing those lines sends the agent straight to the model again. Its docs list current limits: no on-premises build, no SSO or team roles yet, and no paired Codex cost test so far. The Engine plugin installs through the Claude Code plugin marketplace or npm and, per its product page, works in 18 AI coding tools.
Published results (vendor-reported)
Both vendors publish their own measurements with methods attached. They use different tasks, models and metrics, so the numbers below cannot be compared directly, and neither is an independent test.
- Caveman Read less: 33.2% fewer input tokens across whole Claude Code sessions on six file types (885,793 to 591,673 tokens), with 18 of 18 answers correct. The README notes the session run dates from August 2026 on an older engine and that HTML, which has no compressor yet, came out 9.9% worse.
- Caveman Auto: an interim run paused at 85 of 169 paired tasks from SWE-Interact and SWE-Together resolved 43.7% of tasks against 44.3% for Opus 5.5 at xhigh effort, at 0.46 times the cost per attempt. Caveman says the score gap is too small to separate at this sample size.
- Caveman output: its homepage slogan is half the bill, but its own figures are narrower: 65% less readable output on a 10-prompt benchmark, which it says is about 10% fewer total tokens on long coding tasks. Elasticsearch Labs reported 63.6% fewer response tokens for an Elasticsearch adaptation across eight MCP scenarios.
- Context Mode gateway: 168 pre-registered paired Claude Code runs on Opus 5.5 over 10 short tasks cost $28.28 direct and $23.16 through the gateway at list price, with 155 pairs cheaper and all 168 tasks passing in both arms.
- Context Mode long sessions: a long agentic session on Opus 5.5 with a 1M window went from $7.20 to $2.49 (65.5% lower) across 4 pairs, and a 200K-window session with Recall was 37.6% cheaper across 3 pairs.
- Context Mode Engine: the README reports that 315 KB of raw tool output in a session becomes 5.4 KB, a 98% reduction in tool output entering the context, with lower savings on platforms without hooks.
Both vendors warn that fewer tokens do not always mean a lower bill. Caveman notes that cached input, retries and recovery calls affect the result and that a flat subscription may cost the same. Context Mode reports costs at API list price, which is a list-price value rather than cash savings if you use a Claude plan.
Pricing
Caveman's CLI, skill and SDK are free. Caveman Cloud is $0 for one person with monthly allowances of 100,000 events, 600 agent minutes and 25,000 routing decisions, and no card required. Team costs $300 a month with $300 of usage included, no per-seat fees and a 7-day free trial. SSO comes with a $750 a month Scale package, and Enterprise is an annual contract priced on the AI spend Caveman manages. Usage beyond allowances is billed at standard rates, from $0.00035 per event, with 0% markup on tokens.
Context Mode's Engine is free. The gateway is billed by request: Free gives 1,000 requests once, Pro is $49 a month or $490 a year for 50,000 requests a month with 25,000-request packs, Team is priced per seat for two or more seats with 50,000 requests per seat, and Enterprise is a custom yearly contract. Every plan includes every feature, your Anthropic or OpenAI account pays for the model with no token fee added, and when requests run out the agent keeps working by going straight to the model.
Licensing
Caveman's README says the whole repository is Apache-2.0, and its pricing page lists the skill, CLI, SDKs, middleware, extension and Browse as Apache-2.0, with Caveman Code and the standalone cavemem package under MIT and releases before 3.0.0 keeping the licence they shipped with. Context Mode Engine is under the Elastic License 2.0, which is source-available: you can use, modify and redistribute it, but not offer it as a hosted service. The @context-mode/cli package is published under MIT on npm, while the gateway itself is a hosted commercial service.
Beyond token savings
Context Mode's gateway bundles features that go past cost: Memory that keeps facts across sessions and agents, Recall that keeps long sessions from compacting, Search and session resume across machines, Skills loaded from GitHub, Instructions, and Protect, which checks tool calls against rules such as blocking .env reads or pushes to protected branches. Context Mode notes that Protect reads what a call does, so a script file or compiled program can get past it, and it advises keeping the agent's own sandbox on.
Caveman's extras sit around measurement and routing. Caveman Cloud adds traces, cost analytics by agent and model, evals built from your own requests, spend caps and an Autopilot mode that tests fixes on a share of traffic. Caveman's pricing table marks harness optimization as in development, and lists model routing under the agent as coming soon even though the README already documents Auto, so check the current state before relying on it.
Who should pick which
- Pick Caveman if you use several coding agents beyond Claude Code and Codex and want one free tool that runs locally.
- Pick Caveman if you want shorter agent replies as well as smaller tool output, or if you build your own agents and need compression middleware for TypeScript or Python.
- Pick the Context Mode gateway if your team works in Claude Code and Codex and wants managed folding, memory, long-session Recall and a single tool-call policy, and you can accept a hosted service on the request path.
- Pick Context Mode Engine if you want local tool-output sandboxing and session continuity with no account or telemetry, and do not want the model's writing style changed.
Neither vendor documents running Caveman and Context Mode together. Because both change what an agent sends to the model, test any combination on your own tasks before relying on it.
The Bottom Line
Caveman and Context Mode aim at the same problem, coding agents spending tokens on noise, but split along one line: Caveman keeps the work on your machine and adds an optional cloud, while Context Mode's main product is a hosted gateway with memory and policy features and a separate local plugin. Their published savings come from different benchmarks and are vendor-reported, so the fairest test is your own agent, tasks and billing model.