Context Mode sits between your coding agents and the model. Each request costs less, sessions carry what they learned, and your rules decide what runs.
npx @context-mode/cli1,000 requests free, every feature, no card. Supported agents
600 tests, 1 failing17.0 KB → 1.2 KBfoldedcat .envnever reaches the shellblocked2 h laterpicked up, nothing compactedrecalled7 MB registry file407 B back to the modelin codegit push origin mainprotected branchblockeddeploy target?Render, not Hetznerrememberedturn this plan into ticketsto-tickets loadedskillContext Mode acts at three points on each request: before it goes out, before a tool runs, and when a result comes back.
Your agent runs 600 tests. 1 fails.
Sent to the model
Kept in your archive
Your agent reads an API response with 800 orders.
Sent to the model
Kept in your archive
One session, 30 steps. Tokens sent on each call.
Sent to the model
Kept in your archive
Tool output is folded before the model reads it. The full output is archived.
Features Context Saving → Thinking in Code → Active Context → Prompt caching → How saving works →| Fact | Value | Valid from | Until |
|---|---|---|---|
production deployed_to | Hetzner Cloud, Falkenstein | 21:14:38 | 21:15:09 |
production deployed_to | Render, Oregon region | 21:15:09 | now |
The old value closes. Every agent gets the new one.
Older work goes to the archive, word for word.
⚡ context-mode · picked up after a 2 h pause, nothing compacted
⏺ context-mode · skill to-tickets auto-loaded
Matched on name and description. One skill per turn.
Sessions don’t compact. Facts and skills carry across agents and machines.
Features Recall → Memory → Skills → Search → Session resume →.env and .env.*. Examples and templates are allowed.
⛔ context-mode Cage blocked this command, so it was NOT executed:
Blocked by your team’s policy: Read environment files. .env files hold secrets.main, master, production and release/*.
⛔ context-mode Cage blocked this command, so it was NOT executed:
Blocked by your team’s policy: Push to protected branches. Protected branches change through reviewed pull requests.Read through sudo, sh -c and about 130 other wrappers.
The inner host is checked too.
A project folder is fine. Only losses you can't undo are blocked.
Nothing is blocked until you choose rules.
Rules for what agents may run, read and fetch. Nothing is blocked until you choose them.
Features Protect → Rules and presets → Instructions →Claude Code on Opus 5.5. Each task ran direct to Anthropic and through Context Mode, back to back. Cost at API list price.
Every pair counts. No figure for Codex yet. Method, 95% ranges and Haiku 4.5 →
One command signs you in, finds your agents, and shows each change before it writes. Your own model account still pays the model.
It signs you in, asks which agents to connect, and keeps a backup of each file.
npx @context-mode/cliStart a new session in your agent. Work as before.
The console shows tokens, cost, saving and blocks for every request. Open the console →
~/.claude/settings.json
// what the CLI writes, under "env"
"ANTHROPIC_BASE_URL": "https://gateway.context-mode.com",
"ANTHROPIC_CUSTOM_HEADERS": "x-cm-device-key: cmdk_…",
"ENABLE_TOOL_SEARCH": "true"~/.codex/config.toml
model_provider = "context-mode"
[model_providers.context-mode]
name = "context-mode"
base_url = "https://gateway.context-mode.com"
wire_api = "responses"
requires_openai_auth = truenpx @context-mode/cli disconnectQuick start →Your own Anthropic or OpenAI account pays the model, with no fee added on tokens. When your requests run out, nothing breaks. Compare plans →
$01,000 requests, once
About a week of very heavy use. Every feature.
Get started free$20/monthor $200 a year
10,000 requests a month, every feature. For developers who use agents most days.
Start Starter10,000 requests a month. Need more? $2 per 1,000, up to a budget you set.
$49/monthor $490 a year
50,000 requests a month, for developers who run agents all day.
Start Pro50,000 requests a month. Need more? $1 per 1,000, up to a budget you set.
Teams and Enterprise: compare plans.
Protect checks every command, file read, web fetch and tool-server call from Claude Code and Codex before it runs. A blocked call never reaches your shell. In every plan, Free included. How Protect works →
npx @context-mode/cli. Only the base URL changes.
Nothing is blocked until you pick one.
The agent reads why a call was blocked. You see it in the log.
See before you rule.7 run
git push origin mainRunscat .envRunsrm -rf ~Runsterraform destroyRunsaws s3 rm s3://backups --recursiveRunsWebFetch example.comRunsnpm testRunsNothing is blocked. Activity lists the programs, sites and tool servers your agents used, each with Block and Allow.
Where most developers start.4 blocked · 1 logged · 2 run
git push origin mainBlockedPush to protected branchescat .envBlockedRead environment filesrm -rf ~BlockedCore protectionsterraform destroyBlockedDestroy infrastructureaws s3 rm s3://backups --recursiveLoggedChange cloud resourcesWebFetch example.comRunsOther sites allowednpm testRunsA blocked call never reaches your shell: the agent reads the rule and why, and carries on. Logged means it ran, and the log says it would have been blocked.
For regulated teams and unattended agents.6 blocked · 1 run
git push origin mainBlockedPush to protected branchescat .envBlockedRead environment filesrm -rf ~BlockedCore protectionsterraform destroyBlockedDestroy infrastructureaws s3 rm s3://backups --recursiveBlockedChange cloud resourcesWebFetch example.comBlockedOnly sites on your listnpm testRunsOnly the sites and tool servers on your list are reached. Public package registries stay open.
Or write your own rules for sites, commands, scripts, files and tool servers, and test each one before you save it. Rule types →
For you Free · Starter, $20 a month · Pro
For your team Team · Enterprise
Tested in real Claude Code sessions with Claude Opus 5.5 (2026-10-02) and Claude Haiku 4.5 (2026-10-04): every run blocked the push to main and the read of .env, and the work still shipped on a branch. Try it in 5 minutes →
Protect reads the call, not the program: a script file or a compiled program gets past it. Keep your agent’s sandbox on.
Other tools do one of these, on one client or one machine. Sources, dated →
| Context Mode | Built into your agent | Local optimizers | General LLM gateways | |
|---|---|---|---|---|
| One setup for every agent and machine | Yes | One client | One machine | Yes |
| Folds tool output; full output kept | Yes | Summarizes | One machine | No |
| Checks tool calls before they run | Yes | One client | Some commands | By tool name |
| Memory shared by every agent | Yes | No | One machine | No |
| Built for coding agents | Yes | Yes | Yes | No |
Start free: 1,000 requests, every feature, no card. When you run out, nothing breaks.