You’re burning tokens.
ctxpilot stops the bleed.
Your agent re-reads the same files every session. That’s 30-50% of your input tokens — wasted. ctxpilot eliminates the waste for ~$0.001/session.30-50% less spend. 200× fewer tokens per lookup. ~$0.001 per session.
Current beta stack
The problem
You ship the same fix on Tuesday.
You pay for it again on Friday.
Every new session, your AI agent starts from zero. It re-reads package.json for the 47th time. It re-derives the architecture you already explained. It re-discovers the bug you fixed last week. You pay for every byte of that amnesia.
A typical mid-size repo, 30 days of Claude Code use. The same file, parsed and tokenized 47 separate times.
Average ‘fix one bug’ session today. ~65% of that is the agent re-reading files it already understood.
Per developer. Per project. Compounds every time you bring a new agent online.
How it works
Install once. It runs.
The beta installer writes Claude Code hooks, starts a Codex CLI session watcher, boots the local worker, and registers the MCP server. Then you go back to coding. ctxpilot watches, learns, and serves answers when asked.
Hook
Every tool call from Claude Code and OpenAI Codex CLI fires a local event. Nothing leaves your machine.
[hook] tool_use agent: claude | codex tool: edit_file file: auth.ts
Extract
Git diffs, test output, command results — parsed by deterministic rules. Zero AI calls for 90% of memories.
extracted: file_edit 0.91 test_pass 0.86 command_ok 0.93
Store
Content-hashed and stored in local SQLite. Auto-deduplicated. Confidence decays over time.
db: ~/.contextpilot/contextpilot.db hash: e7f3a1 dedup: hit (existing) conf: 0.91 → 0.91
Recall
15 tokens instead of 5,000. Same answer. Agent calls contextpilot_search mid-session and gets the compact, ranked result.
→ "JWT validation added session 42. middleware.ts:47" → 15 tokens · 2 hits
Side-by-side
Same bug fix. Two-thirds fewer tokens.
Left: what you pay today. Right: after npx ctxpilot install.
Without ctxpilot
With ctxpilot
65% fewer tokens. Same result. Every session.This is one bug fix. Multiply by ~80 sessions/month.
The math
Your bill, reduced.
What an active developer on Claude Opus saves in a month — 150 sessions, multi-repo workflow. Heavier usage scales linearly. Run your numbers ↓
can vanish into amnesia today — your agent re-reading the same files. ctxpilot keeps the work the agent did yesterday.
For an active Opus user at 150 sessions/month. That’s $54/mo straight off your Anthropic bill — at $9/mo Pro you net $540/year.
15 tokens to recall a fact from memory. 3,200 tokens to re-read the file. Every. Single. Session.
Projected launch model · replace with your local measured savings after install
Calculator
What you’re actually spending.
Plug in your usage. Real model pricing as of May 2026. Assumes ~75/25 input/output ratio and a projected 42% re-read rate.
Your usage
/ inputsA session = one AI coding task, end-to-end.
Total input + output tokens. Default reflects a typical mid-size repo.
What ctxpilot saves you
/ outputStay on Free. You'll save $4.46/mo at $0 cost. Upgrade to Pro ($9/mo) once you have multiple repos.
Disclaimer: estimates are projections until your local install has measured sessions. Your savings depend on repo size, agent, and how often you start fresh sessions.
Memory engine
Eliminate repeated reads automatically.
Deterministic extraction from git diffs, test output, and commands. No AI calls for 90% of memories. Content-addressed deduplication. Automatic confidence decay. Secret scanning built in.
CLI savings reports
Get a waste receipt from the terminal.
Repeated-read detection, savings proof, project waste reports, and projected monthly savings from local data. No browser app required — your data stays on your machine.
Project waste analyzer
Find expensive context before agents reread it.
Analyze instruction files and project context weight without storing source bodies. The report shows what is likely costing credits every time an agent starts.
MCP integration
75 tokens instead of 15,000. Same answer.
Four MCP tools any compliant agent can call mid-session — Claude Code and OpenAI Codex CLI both supported in v1: contextpilot_search, contextpilot_get, contextpilot_remember, contextpilot_status. The agent gets the result without re-reading five files.
Local-first & private
Your code never leaves your machine.
Memory is stored in a local SQLite database. No cloud sync. No telemetry. No analytics calls. Secret scanner blocks API keys before they’re ever indexed.
- ✓SQLite on your local disk.
- ✓Secret scanner — never indexes API keys.
- ✓No product telemetry in the local worker.
- ✓Uninstall keeps data unless you choose --delete-data.
- ✓Core capture works without cloud sync.
Install
30 seconds. No signup.
Writes hooks into your agent, starts a local worker on port 3848, registers the MCP server, and keeps reports available from the CLI.
Free tier saves you money from session one · no credit card
Pricing
Free already saves you money.
ctxpilot Free pays for itself the moment you install it. Paid plans are in customer beta while licensing, checkout, and support workflows are finalized. Run the numbers ↑
Free
Solo dev · one repo
- 1 project
- 100 memories
- Full token savings
- CLI savings reports
- MCP + CLI
- No card · no trial
Pro
Solo dev · many repos
- Unlimited projects
- 5,000 memories per project
- Waste detection
- Full CLI analytics
- Priority beta support
- License activation
Builder
Heavy use · 150+ sessions/mo
- Everything in Pro
- Unlimited memories
- Advanced waste reports
- Git-aware insights
- Session timelines
- Custom memory rules
Team
Shared memory · whole team
- Everything in Builder
- Shared team memory pool
- Per-seat access controls
- Cross-repo memory policy
- Audit exports
- Dedicated support
Enterprise: self-hosted pilots · compliance review · custom support · contact us
One command from here