context-doctor 🩺

Keep every AI session's context lean, and see exactly what it costs you, without ever making it more expensive.

Long agent sessions fill up with tool output nobody reads again, and a 500k-token session re-reads all of it on every message. The most expensive moment is coming back after lunch: the prompt cache has expired, so the next message re-sends everything at full price. context-doctor measures this on your own sessions and acts on it at the moments that pay.

• Tells you when to /compact, at the moment it pays. When you come back to a large session after the cache expired, the every-prompt hook gives the model the numbers and it offers /compact once. Replayed over 133 days of the author's Claude Code history: 401 such returns, and compacting then would have saved $4,053 net at list price, about $914 a month. It works in every Claude Code surface, including the desktop app. What it saves →
• Compacts earlier, automatically, in every Claude Code surface. Claude Code auto-compacts near the full window (measured: ~970k tokens on a 1M model), so long sessions re-read 400k-900k tokens on every message. context-doctor compact-window replays your history at smaller windows and, if you choose, sets Claude Code's own autoCompactWindow. On the author's last 30 days: a 400k window would have cut input cost 55% ($6,119 to $2,757), for 15 compactions a week instead of 5. A native setting, so it works in the desktop app too. The trade-off →
• Autopilot for terminal and IDE sessions and API apps: a local proxy that clears stale tool output only when the prompt cache is cold, so it never costs more. 9.8% less input cost, no session made more expensive, in the same replay. (The desktop app's Code tab sets its own API address, so autopilot cannot reach it; doctor tells you if that is your case.)
• Counts Claude correctly. Current Claude models pack 2.75 characters per token, not the 4 most tools assume, so estimates built on 4 undercount Claude by about 40%. The ratios were measured from the API's own counts; context-doctor accuracy re-checks them on yours.
• Works where you work: Claude Code (terminal, IDE, desktop app, plugin), Cursor, Codex, Claude Desktop chat, any Anthropic or OpenAI API app, VS Code, CI. macOS, Linux and Windows, Node 20+. Local, keyless, no telemetry, MIT.

Built and maintained by gAI Ventures.

Quick start

See what it would save you first (no install, reads your local Claude Code history):

npx context-doctor savings

That is the author's machine: every session there runs in the desktop app, so the first line is the one that applies. Yours is computed the same way, from your own transcripts: every session replayed request by request through the shipped code, against what you were actually billed. context-doctor savings --share prints a few lines with totals only (no project names or paths) if you want to post your number.

Then install it, whichever way suits you:

# Everything, every app it finds (Claude Code, Cursor, Codex, Claude Desktop)
npm install -g context-doctor
context-doctor install
context-doctor autopilot on     # every new Claude Code session keeps its context lean

# Or as a Claude Code plugin, from inside Claude Code
/plugin marketplace add KushalP1/context-doctor
/plugin install context-doctor@context-doctor

The plugin brings the every-prompt check, the MCP tools, and /context-doctor:savings, /context-doctor:checkup, /context-doctor:autopilot and /context-doctor:compact-window. It needs no npm step: the MCP server ships as one self-contained file, so it also works on Claude Code versions that do not install plugin dependencies.

context-doctor autopilot status shows what autopilot did; context-doctor doctor checks the whole setup. Everything is reversible: context-doctor autopilot off, context-doctor uninstall, or /plugin uninstall.

Where the tokens go
────────────────────────────────────────────────────────
Tool results      ████████████████░░░░░░░░░░░░  57%  ~41k
System prompt     ██████░░░░░░░░░░░░░░░░░░░░░░  21%  ~15k
Assistant replies ████░░░░░░░░░░░░░░░░░░░░░░░░  13%  ~9.4k
User messages     ██░░░░░░░░░░░░░░░░░░░░░░░░░░   9%  ~6.5k

Findings (4)
────────────────────────────────────────────────────────
✖ Message #12 contains a base64/binary blob (~8.2k tokens). [save ~7.4k]
   → Never put base64 in text content — use the provider's file/image APIs.
▲ Tool result at message #7 (web_search) is ~6.1k tokens. [save ~4.9k]
   → Truncate or summarize large tool outputs before they enter history.

What it saves

Measured, not modelled, on every Claude Code session on the author's machine (mostly Opus 5 and Fable 5 with the 1M window), priced the way the prompt cache bills it (cached reads 0.1x, writes 1.25x) at API list prices. context-doctor savings runs the same measurement on yours.

1. Compacting when you come back (every Claude Code surface, desktop app included). 133 days, 401 returns to a session over 150k tokens after more than 65 idle minutes; median context at the return, 517k tokens.

Each return counts only the messages up to the next return, and the compaction request is charged as a cost. Not counted: files the model may re-read after compacting, and what the summary leaves out. That is why this is advice with the numbers attached, not something done for you.

2. Autopilot (terminal and IDE sessions, API apps). 130 days, 43 sessions, replayed through the shipped autopilot code with real timestamps:

The median session saves 1.9%, the best 37.7%: long sessions are where the money is (94% of input cost on that machine came from requests above 200k tokens). These sessions ran in the desktop app, which autopilot cannot reach, so for the author this is what the same sessions would have saved in a terminal; context-doctor savings separates the two for you.

3. Compacting earlier (automatic, every Claude Code surface, desktop app included). The offer in 1 is advice, and on the author's machine it was followed 1 time in 33. Claude Code has a native setting that needs nobody to follow anything: autoCompactWindow in ~/.claude/settings.json makes it compact as if the window were that size. context-doctor compact-window replays each session request by request at several sizes, priced as the cache bills (each simulated compaction pays its own request and a 20k-token summary at the output rate), and changes nothing until you pick one:

29 sessions. The replay's baseline came within 9% of what the transcripts say was billed, on the low side. The cost is real: each compaction replaces the session's history with a summary, so a smaller window trades detail for money. That is why the command shows the table first and sets a window only when you name one (context-doctor compact-window 400k; off undoes it, and a backup of settings.json is kept).

On a Claude subscription you do not pay list price; the same tokens come out of your usage limit instead. Anthropic does not publish how limits weight cached tokens, so read the dollars as the size of the effect, not as your bill.

What happens on each platform

Not claimed, because no process on your machine sends those requests: trimming inside Claude Desktop chat, claude.ai, ChatGPT, Cursor's own subscription models, or Codex signed in with ChatGPT. Those get the rules and the measurements above, not autopilot.

What's new

• 0.25 Compact earlier, by itself: the /compact offer at cold resumes was followed 1 time in 33 on the author's machine, so two changes. The hook now also shows you the notice (it went only to the model, which rarely raised it), with the dollar cost of the message you just sent. And context-doctor compact-window measures and sets Claude Code's own auto-compact window, which works without anyone acting on advice, in every surface including the desktop app: 55% less input cost at 400k on the author's last 30 days. savings shows it as a third lever.
• 0.20 Autopilot: stale tool output cleared from every Claude Code request, only when the prompt cache is cold, so it cannot cost more (measured: 9.8% less input cost, ~$1,080 a month on the author's usage, no session worse); runs as a login service on macOS, Linux and Windows; now also for GPT via OpenAI's Chat Completions and Responses APIs.
• 0.24 A full audit, fixed: the session parser now rebuilds API messages from Claude Code's one-row-per-block transcripts (it had split each reply into several messages, breaking tool pairing in exported conversations and inflating profiles); the proxy forwards every non-conversation endpoint and never re-encodes bodies it does not rewrite (model listing and file uploads failed or were corrupted through it); report and the dashboard no longer count compaction as tokens context-doctor saved, and MCP sketches no longer inflate them; watch follows plain conversation files; MCP inputs are bounded. 0.23: four dependency vulnerabilities patched, no false /compact alarm after compacting, savings --share.
• 0.22 Honest about where each lever works, and a new one that works everywhere: the every-prompt hook now offers /compact when you return to a large session after the cache expired (measured: ~$914 a month on the author's history), including in the desktop app. Found and fixed: autopilot cannot reach the desktop app's Code tab, and doctor now says so instead of reporting ✓; savings counted whole-file history as "last 30 days"; the hook quoted per-message cost at the uncached rate (10x too high). --version added.
• 0.21 See it before you install it: npx context-doctor savings replays your own recent sessions through autopilot and shows what it would have saved, against what you were actually billed. Install as a Claude Code plugin from inside Claude Code. Ready for the official MCP Registry (io.github.KushalP1/context-doctor), which every release now publishes to; any MCP client can launch it as npx -y context-doctor mcp.
• 0.20.1 Releases that finish themselves: one tag publishes to npm and creates a GitHub release with the Claude Desktop bundle (signed when a certificate is configured); the editor extension is ready for the VS Code Marketplace and Open VSX.
• 0.19 Measured Claude tokenizer: estimates were 40% low for Claude; fixed from the API's own counts, with a per-model check in accuracy.
• 0.18 proxy --token for putting the proxy on a public URL safely. 0.17 Claude Desktop: a profile_context the model can afford to call from chat, .mcpb bundle, standing preferences for web and mobile. 0.16 Codex. 0.15 Cursor.

Full history with the measurements behind each change: ROADMAP.md.

Setup details

install configures every app it detects and does not stop at the first problem: a corrupt Claude Desktop config still gets you Claude Code and Cursor. It does not pretend either. Any target that failed is named with a ✗ line, the summary reads "Done with N problem(s)" instead of "Done.", and the exit code is 1, so dotfiles and onboarding scripts can react. A broken config file is never overwritten; fix it and re-run.

npx context-doctor install works too, but autopilot needs the global install: a background service cannot point into npx's cache, which npm deletes at will.

No API keys, ever. Everything is deterministic local code; when an LLM is needed (summarizing pruned history), the model already running in your app does it. The proxy forwards your app's credentials untouched — context-doctor itself holds nothing.

What install actually does — and what happens in every session after

One run of context-doctor install writes these (each config edit makes a .backup first; uninstall reverses all of it):

1. Claude Desktop config (claude_desktop_config.json) — registers the MCP server
2. Claude Code config (~/.claude.json) — registers the MCP server
3. Cursor config (~/.cursor/mcp.json) — registers the MCP server
4. Agent Skill → ~/.claude/skills/context-doctor/ — context-hygiene playbook for Claude Code
5. Every-prompt hook → ~/.claude/settings.json — the per-query context check for Claude Code (Cursor runs it too)
6. Codex, when present: MCP server in ~/.codex/config.toml, the hook in ~/.codex/hooks.json, the skill in ~/.codex/skills/
7. With --statusline: live context size, cache share and cost in Claude Code's status bar

context-doctor autopilot on is separate and opt-in: it adds the background proxy service and one line (env.ANTHROPIC_BASE_URL) to ~/.claude/settings.json, after the proxy has answered a health check.

In every chat afterward (Claude Desktop, Cursor): when the conversation starts, the app launches the MCP server, which hands the model standing instructions that stay in force for the whole chat:

• summarize large pastes and tool results instead of carrying them verbatim,
• reference earlier content instead of re-quoting it, never inline base64,
• once the chat passes ~30 turns or accumulates big pastes, proactively offer to profile it,
• answer any "what's eating my context / cost / latency" question by calling profile_context, not by guessing.

In every Claude Code / Cowork session afterward: all of the above via MCP, plus two more layers:

• the skill loads whenever context work is relevant, and
• the hook runs on every single prompt you send: lean sessions cost a ~1ms file-size check; once a session is heavy it profiles on growth events and injects a note the model sees with your message — actual token count, cost per message, the single largest recoverable waste — with instructions to work leaner and offer you compaction. It re-fires only after ~40% further growth, can never break a prompt (any failure exits silently), and logs each deep check to a small local ledger that feeds context-doctor report.

What it never does: delete or rewrite your history without asking (pruning is consent-only, and the model writes the replacement summary so nothing is lost silently), send data anywhere (everything runs on your machine), or touch an API key.

Do you need MCP? Only sometimes — all the ways to use context-doctor

MCP is just one of seven delivery mechanisms. It's only required when you want the AI inside a chat app to run the tools itself. Everything else works without it:

Practical upshot: a developer who only wants cheaper, faster API calls never touches MCP (proxy + CLI). A Claude Code user gets the hook and skill without MCP either — the MCP server just adds in-chat tools on top. install sets up all of it at once precisely so you don't have to think about which mechanism is which.

All commands at a glance

What "always-on" means, per surface

Without autopilot nothing runs in the background: the hook, skill, MCP server and its instructions are delivered by the apps themselves at the right moment. With autopilot, one small proxy runs as a login service (launchd / systemd user service / logon task) and is restarted by the service or by the next prompt's hook if it ever stops.

Or use the CLI directly, no install needed:

npx context-doctor analyze conversation.json --model claude-sonnet-5

npx context-doctor optimize conversation.json --out slimmed.json

Input is any of: OpenAI chat format, Anthropic messages format (with system and content blocks), or a bare [{role, content}] array. Use - to pipe from stdin.

Reports include dollar and latency estimates, not just tokens:

Cost:  ~$3.30 input per call · ~$3302 per 1k calls · ~13.2s of latency per call (estimates)
...
Potential recovery: ~5.9k tokens (~73% of context) ≈ $17.69 per 1k calls, 0.2s faster per call

Profile your actual Claude Code sessions

npx context-doctor session            # profile your most recent session
npx context-doctor session --list    # browse sessions

Parses the transcripts Claude Code writes locally and answers "where did my tokens go today?" — it will happily tell you that one giant skill load is 67% of your context.

Autopilot: every Claude Code session keeps its own context lean

context-doctor autopilot on        # once; survives reboots
context-doctor autopilot status    # what it has done
context-doctor autopilot pause     # instant passthrough, nothing restarts
context-doctor autopilot off       # remove it

autopilot on runs the local proxy as a background service (launchd on macOS, a systemd user service on Linux, a logon task on Windows), waits until it answers, and only then points Claude Code at it through env.ANTHROPIC_BASE_URL in ~/.claude/settings.json. Every Claude Code session started afterwards in a terminal or an IDE sends its requests through it. Not the desktop app's Code tab: it runs Claude Code "host-orchestrated", sets ANTHROPIC_BASE_URL from its own account configuration, and drops the same key from settings files (checked in Claude Code 2.1.284). autopilot on, autopilot status and doctor look at which surface your recent sessions ran in and warn when autopilot cannot reach them; for desktop sessions the levers that work are the cold-resume advice and compact-window. Your login (subscription or API key) passes through untouched.

What it does to each request: once old tool output adds up to 20k+ tokens (file reads, shell output, search and web results, screenshots inside them), it replaces that output with a one-line note, keeping the 3 most recent results. The tool call stays in the history, so the model can simply run it again if it needs the output. Your messages, its answers, answers you gave to its questions, subagent reports and MCP results are never touched.

Why it cannot make a session more expensive. A prompt cache matches a byte-identical prefix, and changing old history re-bills everything after the change at the write rate (1.25x instead of 0.1x). So autopilot changes history only when the cache is cold anyway: after an idle gap longer than the cache lifetime the request itself declares (1 hour for Claude Code on a subscription, 5 minutes otherwise), when the whole prompt is re-written regardless. Once cleared, an output stays cleared on every later request, so the prefix is identical between clearings and the cache keeps hitting.

Measured before it shipped, by replaying every Claude Code session on the author's machine (43 sessions, 130 days) request by request through the shipped code, priced as the cache bills it, with real timestamps: 3.8 billion input tokens not sent and $4,694 at list price, ~$1,080 a month (details). scripts/replay-autopilot.mjs does this on yours:

94% of input cost on that machine came from requests above 200k tokens, which is where the clearing lands. Smaller requests also mean later auto-compaction and a faster first token.

Never in the way: anything it cannot parse is forwarded unchanged; only Anthropic /v1/messages requests are touched. It adds about 7 ms to a 2.9 MB (~1M token) request. If the proxy dies, the service restarts it within seconds, and the every-prompt hook checks it before each prompt and starts it if needed (measured: 0.6 s once, when it had to). pause turns it into a passthrough without restarting anything; off removes the setting before stopping the service, and sessions started while it was on need a restart.

What it cannot reach (no process on your machine sends those requests): Claude Desktop's chat tab and claude.ai, which use the standing instructions and the profile_context sketch instead; Cursor's own models (Cursor's servers call the model; with your own OpenAI key, see the tunnel section below); and Codex signed in with ChatGPT, where the every-prompt hook still reports context size.

Always-on: optimize every request automatically

Run the proxy and every Anthropic/OpenAI API call your apps make gets optimized in flight — no code changes:

npx context-doctor proxy

Then point your app or SDK at it:

export ANTHROPIC_BASE_URL=http://localhost:8787      # Anthropic SDKs / tools
export OPENAI_BASE_URL=http://localhost:8787/v1      # OpenAI SDKs / tools

The proxy dedupes repeated content, trims stale tool results, and strips base64 blobs from the message history of each request, then forwards it to the real API. Your API key passes through in headers untouched, streaming (SSE) works unchanged, and per-request savings are logged with upstream latency:

[context-doctor] POST /v1/messages → 200 in 842ms | optimized 7.3k → 518 tokens (2 changes) | session total: 6.9k tokens ≈ $0.021 saved

Only conversation requests (/v1/messages, /v1/chat/completions, /v1/responses) are ever rewritten. Every other endpoint a client uses through the proxy (listing models, uploading files, counting tokens) is forwarded untouched, to Anthropic when the request carries anthropic-version or x-api-key and to OpenAI otherwise, and request bodies the proxy does not rewrite reach upstream byte for byte.

GET http://localhost:8787/stats returns cumulative savings (requests, tokens, estimated USD), exact upstream usage read from every response (JSON and SSE), and prompt-cache advisories — the proxy watches your real traffic and flags big stable prefixes missing cache_control or prefix churn that silently re-bills the cache. Per-model behavior via --config: The advisor now says where: for a large system/tools prefix it names the block to mark (the last tool definition, or the last system block), and when the older messages were byte-identical to the previous request it names the exact message to put cache_control on, with the token count that run is re-billing each turn. Anyone who already placed breakpoints is left alone.

{ "routes": [{ "modelPrefix": "gpt", "strategies": ["strip-base64"], "keepRecent": 4 }] }

Because prompt caching matches byte-identical prefixes, deterministic strategies are chosen so repeated requests stay stable — but if you rely on aggressive cache prefixes, start with --strategy strip-base64 --strategy dedupe and add more as you verify.

Putting the proxy on a public URL (Cursor with your own OpenAI key, remote apps)

Some apps let you set a base URL but call it from their servers, not your machine. Cursor is one: with your own OpenAI key, "Override OpenAI Base URL" is sent to Cursor's backend inside the model configuration, and Cursor's servers make the request (only the key-verification ping is client-side; we checked the app bundle, 3.18.25). So 127.0.0.1 cannot work there; the proxy has to be reachable from the internet, and an open relay on the internet is a bad idea. Hence the token:

npx context-doctor proxy --token "$(openssl rand -hex 16)"   # or CONTEXT_DOCTOR_PROXY_TOKEN=...
ngrok http 8787                                               # or any HTTPS tunnel / reverse proxy

With --token, every path except /health must start with /t/<token>/; anything else gets 401 before any upstream call, and the comparison is constant time. Then in Cursor: Settings > Models > OpenAI API Key > Override OpenAI Base URL = https://<your-host>/t/<token>/v1. Every agent request Cursor makes with your key now passes through the proxy: deduped, stale tool results trimmed, base64 stripped, real usage counted in /t/<token>/stats. This is the one Cursor path that is model-independent and needs no hook. It applies only to BYO-key traffic; Cursor's own subscription models never leave Cursor's servers.

Your API key still rides in the request headers, as before. The token protects the relay, not the key; keep the tunnel HTTPS.

Note on desktop chat apps: Claude Desktop and the ChatGPT app talk to their own backends — no tool can sit in that path. For those, use the MCP integration below (Claude Desktop gets the standing rules and a cheap profile_context sketch call) and context-doctor instructions --copy for the per-account preferences.

Use with the Claude & ChatGPT apps

context-doctor ships an MCP server, so the AI itself can profile and slim context on demand. Before the setup table, the honest question: on which surfaces does it act by itself, and on which does it only nudge? MCP gives a server no way to see the conversation or intercept a turn; only a hook or a place in the data path can do that.

So "every chat inherently better" is true for Claude Code, Cursor, Codex and the proxy; for Claude Desktop it is "the rules ride in every chat and the checkup is one cheap tool call away"; and not a claim we make for the ChatGPT chat UI.

Where there is no hook and no MCP at all (claude.ai on the web, the Claude and ChatGPT phone apps, plain ChatGPT): the app's per-account preferences are read on every turn, which is the closest those surfaces have to a hook. context-doctor instructions --copy puts the ~180-token rules on your clipboard and tells you where to paste them (claude.ai Settings > Profile; ChatGPT Settings > Personalization > Custom instructions).

Do you need to configure anything by hand? Usually no:

For any other MCP client, the server entry is:

{
  "mcpServers": {
    "context-doctor": {
      "command": "npx",
      "args": ["-y", "context-doctor-mcp"]
    }
  }
}

How it works in Claude Desktop, step by step

1. Run npx context-doctor install (writes the config above for you) and restart Claude Desktop. Or open the .mcpb from the latest release: same server, no npm, installs as an Extension.
2. From then on, every conversation carries context-doctor's standing instructions. The MCP server hands them to Desktop on connect and Desktop puts them in front of Claude: summarize big pastes instead of re-quoting them, refer to earlier content by name, never inline base64, and past ~30 turns or on any question about tokens, cost, speed or limits, call profile_context before answering.
3. That call is cheap on purpose. Claude cannot export a Desktop chat, so it passes a sketch: how many turns, which blocks are large, repeated, stale or images, with one size hint each (~120 tokens). The server sizes it (usually within ±20%, measured; see "Why token counts are ~"), prices the per-turn re-read (on a subscription that is what spends your usage limit), and returns ranked findings with the action for each: "summarize the nginx config into the points still needed", "refer to test output by name", "offer a 300-token handoff summary for a fresh chat". The reply ends with an instruction to apply the top one, not just suggest it.
4. One click instead of asking: the context_checkup prompt in the + menu sends that request for you.
5. Say "optimize it" on an exported conversation and Claude applies the safe fixes; if you agree to pruning old history, Claude itself writes the replacement summary (that's the no-API-key summarization).
6. For the same rules on your phone and on claude.ai, where no MCP server runs: context-doctor instructions --copy, then paste into Settings > Profile > personal preferences.

What this does not do: read the chat behind Claude's back or trim it for you. Desktop chat has no hook API and no transcript on disk (checked in the app bundle, v2.2553). The model does the trimming, when the rules and the checkup tell it to.

How it works in ChatGPT, step by step (honest version)

ChatGPT's MCP support differs fundamentally from Claude Desktop's: it never spawns local processes. Its custom connectors (developer mode) have OpenAI's servers connect to a URL — so the MCP server must be reachable from the internet.

1. Normal ChatGPT (no developer mode): no MCP at all. context-doctor still helps via the CLI: export the conversation and run npx context-doctor analyze chat.json --model gpt-5 / optimize — no account settings required.
2. ChatGPT developer mode: run our HTTP transport somewhere reachable — context-doctor-mcp --http --port 8808 on a small host (bind --host 0.0.0.0 there), or expose your machine temporarily with a tunnel (ngrok http 8808). Then Settings → Connectors → Advanced → Developer mode → add connector with URL https://<your-host>/mcp.
3. Once connected, GPT gets the same three tools with the same trigger guidance: ask "what's eating my context?" → it calls profile_context; "optimize it" works the same, including GPT writing the pruning summary itself.

Security note for step 2: the HTTP endpoint is unauthenticated — put it behind your tunnel's auth or a reverse proxy if it stays up long-term.

claude.ai on the web and the phone apps

Your local MCP server can't reach the website, but the behavior can. Two options: context-doctor instructions --copy and paste into Settings > Profile > personal preferences (applies everywhere you are signed in, phone included), or upload skills/context-doctor/SKILL.md under Settings → Capabilities → Skills. Either way web and mobile conversations gain the same standing habits: summarize, don't re-quote, offer a handoff when heavy.

MCP tools

Use it as a library

import { parseConversation, profileConversation, optimizeConversation } from "context-doctor";

const profile = profileConversation(parseConversation(chatJson), "claude-sonnet-5");
console.log(profile.totalTokens, profile.findings);

const { conversation, tokensBefore, tokensAfter } = optimizeConversation(chatJson, {
  strategies: ["dedupe", "trim-tool-results", "strip-base64"],
});

Subagents: their own windows, your bill

A subagent has its own context window, so its tokens are correctly absent from the parent's profile. They are not absent from the bill. Claude Code writes each one to <session>/subagents/agent-<id>.jsonl, and session now reads them:

Subagents (their own windows, your bill)
────────────────────────────────────────────────────────
49 subagent(s) made 4560 API calls: 626.0M input billed, 1.0M output, ~$608.51.
That is 10% on top of the parent session's own input cost ($5846.45), and none of it appears in the profile above.
    $42.52   51 calls  ctx   319k    6m  You are auditing part of a FastAPI backend at /Users/kp/tech
    $37.48  107 calls  ctx   229k    6m  You are auditing the Turtle AI backend (FastAPI, Python) at
  … and 44 more
18 subagent(s) ended above 200k tokens of context. A subagent that big is doing a main session's job; give it a narrower brief, or split the task.

Per subagent: what it was asked, how many calls it made, the context it ended with, how long it ran, and its cost at list price with cache reads and writes priced correctly. Models without a price on file are counted but marked unpriced rather than costed at zero. On this machine that was 195 subagents across 19 sessions and about $1,844 at list price that no profile had ever shown.

The same number inside VS Code and Cursor

An extension in vscode/ puts context health in the editor's own status bar:

⌁ ctx 848k · 85% · cache 100% ⚠

It reads the newest Claude Code transcript for the open workspace folder, shows live context, share of window and cache share, turns to the warning colour past 70% (configurable), and opens a terminal running context-doctor session when clicked. Nothing leaves the machine; it only reads files Claude Code already writes. Until it is on the marketplace, build and install it locally:

cd vscode && npm ci && npm run package
code --install-extension context-doctor-vscode-0.1.0.vsix     # or: cursor --install-extension …

Context health in Claude Code's status bar

context-doctor install --statusline

Claude Code shows the first line a statusLine command prints, on every refresh, while you type. With this on, that line is the number that matters:

ctx 801k/1.0M ▮▮▮▮▮▮▮▮░░ 80% ⚠ · cache 100% · $12.34

Live context against the model's window, a warning mark from 70%, the share served from cache, and the session's cost. It reads the size from the status payload when Claude Code provides it, and otherwise from the last 256KB of the transcript (about 1ms on a 20MB file; 80ms end to end including Node startup). It is opt-in and polite: there is only one status line, so it never overwrites one you already have, and uninstall removes only its own. Any failure prints nothing rather than an error.

This is the "editor status bar" roadmap item, delivered for the editor most users of this tool are actually in. A VS Code / Cursor extension for the same number remains future work.

Does a smaller context actually help? Measure it

Everything else in this tool measures what is in the context. None of it can tell you whether the task succeeded, so a smaller transcript can be a cheaper failure. experiment is the honest test:

context-doctor experiment \
  --task "Add a null check to parseHeader in src/parse.ts and make the tests pass" \
  --check "npm test" \
  --existing 9c6f9dc9-457b-4d09-bf6d-a499c2f2f919 \
  --model sonnet --budget 1

It runs the task in a fresh session, resets the tree to the starting commit, then runs it again forked from your existing session (--resume … --fork-session, so your real session is never touched), same model, same tools. For each arm it records what you were billed (input, cache read, cache write, output), cost, wall clock, turns, and whether your --check command passed, then puts them side by side:

                               fresh      existing
──────────────────────────────────────────────────
billed input                     20k           65k
  of which cache read              0           60k
  of which cache write          8.0k          4.0k
cost                          $0.110        $0.420
wall clock                        4s            9s
check                           PASS      FAIL (1)

Verdict: fresh was cheaper AND passed; existing failed the check. Clear win for fresh.

The verdict line is the point: cheaper only counts if it also passed. Because this spends your Claude budget it caps spend per arm (--budget, default $1), refuses to start on a dirty tree (both arms must begin from one commit, and the tree is reset between them), and refuses to run from inside a Claude Code session, where the CLI cannot start. --dry-run shows the exact commands first.

Exact counts, and what they teach the estimator

The default token count is a chars-per-token heuristic so everything runs with no key and no tokenizer, with ratios per provider (see "Why token counts are ~" below). analyze --exact fetches a true count for the exact bytes just estimated (Anthropic's count-tokens API with ANTHROPIC_API_KEY; tiktoken for GPT if installed) and prints the drift.

Since 0.13.9 it also remembers the comparison, per model family, on this machine, and later estimates for that family are scaled by it. Nothing about this is silent: the profile header says estimates calibrated +12% from 3 exact count(s) you ran on this machine. No exact count ever run means no calibration and unchanged numbers; out-of-range samples are ignored; CONTEXT_DOCTOR_NO_CALIBRATION=1 returns to the raw heuristic. Samples are tied to the heuristic they were taken against: after 0.19 changed Claude's ratios, older samples are ignored and learning restarts, rather than stacking an old correction on a fixed estimator.

What it detects

• Oversized tool results — the #1 context killer in agent loops
• Oversized tool calls — a Write or a cat > file <<EOF puts the whole file in context permanently. In file-heavy agent work these outweigh every tool result combined, and --strategy trim-tool-calls reclaims them
• Duplicate content — the same doc/result pasted twice
• Near-duplicates — the same doc re-pasted with different surrounding words (shingle similarity, ≥60%)
• Repeated file reads — the same file pulled in three or more times, every copy still in context. Counts shell reads too (cat, head, tail, less), which is where most of them hide in agent sessions
• Retained error output — stack traces and failed commands kept verbatim long after the fix landed
• Repeated identical tool calls, split into the two things they can mean: a retry (the same call after a failure, where the fix is in the error text, and three or more is a loop) and a re-read (the same call after a success, where the model forgot it already had the answer). Across 42 local sessions that was 15 retries against 151 re-reads, so the old combined advice was wrong for most of them
• Base64 / binary blobs in text content — checked by character distribution, not just alphabet, so hex digests and long identifiers are not mistaken for encoded binary
• Long history past the point where models track the middle
• Cache-hostile ordering — volatile content before stable content breaks prompt caching (Anthropic cache_control, OpenAI automatic prefix caching)
• Window pressure — usage % against the target model's real context window

What it fixes (deterministically — no LLM calls, no API keys)

trim-tool-calls is the big one for agent sessions. Writing a file through a tool call puts the entire file in context permanently, so in file-heavy work the calls outweigh every tool result combined — on a real 278k-token session, the default set reached 248k and adding trim-tool-calls reached 102k. It is opt-in because it edits what the model itself wrote, and that is measured, not cautious: across 42 real sessions, of 73 large writes 18 were later edited and 16 of those edits had no read in between, meaning the model built the edit from its own earlier Write input, at a median distance of 43 messages. Trimming those would turn each into a failed edit plus a recovery read. In the offline optimizer the rest of the conversation is known, so exactly those writes are left intact and the 63% never touched again are trimmed; in the live proxy the future is not known, which is why it stays off there unless you turn it on.

Optimization is cache-aware. Prompt caches match a byte-identical prefix, so editing a message in the middle invalidates everything after it — and the naive "trim everything older than the last N messages" boundary moves every single turn. On a 25-turn agent conversation that invalidated the cached prefix on 22 of 24 turns, paying the 1.25x cache-write price on the whole prefix to save a few hundred tokens. The trim boundary is quantized so it holds still between steps (8 of 24 on the same fixture), while still reaching 15 of 20 tool results.

The step size is a trade-off, not a formula, and it is yours to set: "trimBoundaryStep": 20 in .contextdoctorrc (default 10). Measured on a growing agent session at 400 turns: a step of 10 invalidated the cache on 21% of turns with ~2 stale results waiting on average; 20 gave 12% and ~4; 40 gave 10% and ~9. Adaptive steps were worse everywhere, because a step that changes size moves the boundary by itself. Heavy API users who lean on caching want a bigger step; interactive users who want stale output gone promptly want a smaller one.

Everything the optimizer does is inspectable: it prints exactly which messages changed and how many tokens each change saved.

Summarization without an API key: when prune-history runs through the MCP tools, context-doctor hands a digest of the pruned turns back to the model that called it (the Claude/GPT already running in your app) and asks it to write the replacement summary — LLM-quality compaction, zero extra cost, no keys.

The Agent Skill

skills/context-doctor/SKILL.md (installed by npx context-doctor install) teaches Claude to practice context hygiene proactively: summarize big tool results after consuming them, never re-paste duplicated content, keep stable content cache-friendly, and offer compaction when a session gets heavy. In Claude Code and Cursor the every-prompt hook enforces the heavy-session part; the skill covers the habits in between.

Measuring the impact: context-doctor report

npx context-doctor report

One report for your whole machine, led by a headline of tokens context-doctor saved, built only from measured sources:

• exact proxy savings (real before/after on every request),
• exact savings from every optimization applied via the CLI or the in-chat tools — split by model family (Claude vs GPT), with dollar estimates,
• and, on its own line, not in the total: sessions that got smaller after a hygiene warning. That is mostly compaction, by you or by Claude Code, and a warning may or may not be why it happened, so it is not credited to context-doctor (until 0.24 it was, which overclaimed). The dashboard keeps the same split.

Honest measurement note: proxy numbers are exact. Session numbers are measured-now. What no tool can report is the counterfactual — tokens Claude avoided adding because of the hygiene guidance — since the same session can't be re-run without it. The report says so instead of inventing a number.

Context budgets (.contextdoctorrc)

Drop a .contextdoctorrc in a project (or your home directory) and context-doctor enforces your limits instead of its defaults:

{
  "budget": { "maxTokens": 120000, "maxCostPerMessageUsd": 0.5, "maxWindowPct": 60 },
  "strategies": ["dedupe", "trim-tool-results"],
  "keepRecent": 6
}

The nearest file wins (walking up from the working directory, then ~). analyze and session print a budget verdict, the every-prompt hook uses maxTokens as its warning threshold and names the breach to the model, and optimize/proxy pick up the defaults when you do not pass flags.

Prompt-cache economics (Claude Code sessions)

Caching is the largest lever on LLM cost, and transcripts record exactly how it went — so session reports it as fact rather than estimate:

Prompt cache: 95.6% of input served from cache across 1117 requests
  read 558.6M · written 25.5M · uncached 2k
  input cost $438.88 — caching saved $2481.89 against $2920.76 uncached (list prices)

A cache read bills at ~10% of input while a write bills at ~125%, so a session that keeps invalidating its prefix can cost more than one with no caching at all. context-doctor warns on the two failure modes: a low hit rate (something early in the prompt changes every request) and cache churn (writes rivalling reads).

Enforce a budget in CI

npx context-doctor analyze conversation.json --fail-over-budget

Exits 1 when the .contextdoctorrc budget is breached, so a pull request can be gated on context size the same way it is gated on tests.

Sharing a profile safely

A profile quotes message previews and file paths, so pasting one into an issue pastes fragments of real work. --redact keeps every number and the finding structure but replaces content with [redacted] and masks paths:

npx context-doctor session --redact

context-doctor doctor is safe to paste as-is: it reports integration status, never conversation content.

Performance: what context-doctor itself costs

A tool that promises speed must be near-free. Measured overhead per touchpoint:

Net effect is strongly negative overhead: the tokens these touchpoints save on every subsequent call dwarf what they cost.

Why token counts are "~" (and where they are exact)

Counting exactly needs each provider's tokenizer, so the default is a chars-per-token heuristic, with ratios per provider and denser ones for code and JSON. It keeps the tool offline and zero-config.

Which row applies: the model you pass, else the request's own model field, else the request's shape (Anthropic's system field or tool_use blocks mean Claude). Cursor transcripts record no model and use Anthropic-style blocks, so Cursor sessions are counted at Claude density; for a GPT model in Cursor that reads about 40% high.

How the Claude figures were measured, with no key. Claude Code transcripts record what the API billed, and two things in them are exact. A reply with no thinking block is billed as exactly its output_tokens, and all of it is visible text: 504 replies gave a median of 2.75 chars/token (p10 2.4, p90 3.0). Between two consecutive API calls the prompt grows by exactly what was appended; when that is one large block, its size is the growth minus the previous reply: 474 blocks of code and tool output gave 2.4 (p10 2.1, p90 2.8). The ratios this tool used until 0.19 (4.0 / 3.2 for everything) undercounted current Claude models by about 40%: hook warnings came late, savings and costs read low, and the proxy stayed silent on cacheable prefixes between 1,024 and ~1,670 tokens. context-doctor accuracy re-runs both measurements on your own sessions and prints them per model beside the ratio in use, so the next tokenizer change shows up as a number, not a surprise. On this machine every model lands within ±9%.

Two ways to get real numbers instead of estimates:

• analyze --exact uses the Anthropic count-tokens API for Claude models (set ANTHROPIC_API_KEY; opt-in network call, key never stored) or tiktoken for GPT models (install it alongside), and reports how far the heuristic drifted.
• Sessions report measured tokens automatically. Claude Code transcripts record what the API actually charged, so session, the hook and the reports use that figure when it is present — no key, no estimate.

One honest caveat worth knowing: a transcript stores the conversation, not the harness's system prompt, tool schemas or skills (about 54k tokens before the first turn in Claude Code here), nor the reminders it injects each turn. With the corrected ratios the transcript accounts for a median 56% of each turn's billed growth; before 0.19 this read 39%, and about a third of that "invisible" gap was the estimator. That is why sessions prefer the reported figure, and why the message breakdown is labelled as covering messages only.

The chat-app sketch (Claude Desktop, see above) is coarser by design, because the model describes the chat instead of sending it. Its sizes are measured, not assumed: a plain exchange is 2,060 chars (median of 1,283), a code line 42 chars (719 source files), a log line 56 (2,095 tool outputs), a word 6.3, all converted with the model's own ratio. Measured error: the total for a chat of 30+ exchanges from its turn count alone is within -20% to +9% (p10 to p90); a code block sized by lines is within about ±25%, by chars about ±15%. Logs vary from 38 to 100 chars a line, so the tool asks for their size in chars.

Roadmap

See ROADMAP.md for the full plan with rationale. Headlines: v0.5 trust & automation (tag-based publishing, doctor self-check, live watch), v0.6 accuracy (exact tokenizers, semantic dedupe, more session formats), v0.7 proxy pro (response accounting, prompt-cache advisor), v1.0 budgets + local dashboard. Non-goals, permanently: cloud services, telemetry, silent history rewriting, mandatory API keys.

Contributions welcome — this project is small on purpose. Open an issue before a big PR.

Releasing (maintainers)

One-time setup. Each row switches on one channel; any that is missing is skipped with a notice, never a failed run. Secrets go in the repo's Settings > Secrets and variables > Actions:

Each release:

npm version minor        # or patch/major: bumps package.json and tags vX.Y.Z
npm test                 # the full suite; CI runs it on 3 OSes x Node 20/22/24
git push --follow-tags   # npm publish + GitHub release with the .mcpb attached

For the editor extension: bump vscode/package.json, add a vscode/CHANGELOG.md entry, then git tag vscode-vX.Y.Z && git push --tags. The .vsix is attached to a GitHub release either way.

Locally, npm run build:mcpb builds the bundle (MCPB_SELF_SIGNED=1 exercises the signing path with a throwaway certificate; Desktop still warns for those).

What the npm download number measures. install writes npx -y context-doctor-mcp into MCP configs, and npx re-fetches the tarball whenever a new version exists. So every release is downloaded once by every active install within about a day, and the daily count is almost entirely those refreshes: on this package, release days run ~170 downloads and non-release days ~27. Read it as "size of the active installed base × number of releases", not as new users — a quiet week with no releases will look like a decline while nothing has changed. Two corollaries: the release-day figure is a live count of machines running context-doctor, and a broken release reaches all of them automatically, which is why prepublishOnly runs the full test suite. npm's stats also lag by several days and occasionally record a day as zero; a zero on a release day is a gap in their pipeline, not in usage.

Known gotcha: if npm publish fails with 404 Not Found - PUT …/context-doctor on a package that clearly exists, the real cause is an expired npm login token — npm reports unauthenticated publishes as a 404, not a 401. Check with npm whoami; if that errors, run npm login and publish again.

Also keep the MCP server version in src/mcp.ts in sync with package.json, and remember dist/ is committed — run npm run build before committing so the CI dist-sync check passes.

Contributing

Issues and PRs welcome — see CONTRIBUTING.md for the six rules that keep this tool trustworthy (no API keys, nothing leaves the machine, no silent data loss, measurements not guesses, the hot path stays cheap, tests with every change) and a list of good first issues. What is planned next lives in ROADMAP.md.

License

MIT © gAI Ventures