Ellie CLI
One binary that sits next to Claude Code, Codex, Cursor and Kiro. Ellie reviews your working copy, shows where your agent’s tokens went, and answers for incidents, security and your team’s docs. Your agent can run it too.
$ curl -fsSL https:/$HOME
GitHub Copilot
4/6 set up /startFront ends
A command never exists in one front end and not the others. Piped, every screen prints as one final frame.
ellie◉‿◉ › /█arrows to move
A fixed frame you keep open. Type / for the command palette, and shift+tab to switch who answers.
ellie --inline~/src/insights $ ellie --inline
◉‿◉ › review
◆ reviewed 0:41 · 3 files · found 3
● Day boundary computed in local time
● Unbounded fan-out to Slack
● Return shape changed without updating
◉‿◉ › /quit
~/src/insights $ █
Prints into your real scrollback and never takes over the terminal. Suited to tmux panes and narrow terminals.
ellie --panel4 budgetj/k to cycle
Number keys switch screens, with nothing to type. Made for a second monitor.
For agents
Everything Ellie prints, an agent can read. The binary also installs skills and an MCP server.
$ ellie incident show EKU-a3f82b1c --json | jq
{ "eku_id": "EKU-a3f82b1c", "root_cause": "refresh skipped when the token expires mid-request", "recommended_fix": "retry once after refreshing the token"}
Incident and security commands print --json, so an agent parses fields instead of scraping a table.
Installed into the agent’s own skills folder, so Claude Code can call them by name.
◉‿◉ ›fix the worst payments incident this week
enter approve esc cancel
/agent action turns a request into a plan of Ellie commands. You approve it before anything runs.
Where the tokens go
Your agents already keep session logs on disk. Ellie reads the last 28 days of them and sorts the avoidable spend into five classes, each with the skill that removes it.
18¢
of every dollar spent on AI coding tools becomes shipped product. The other 82¢ goes to rework and maintenance.
Token Maxxing Report
Entelligence Research · May 2026 · 1M+ pull requests
% of avoidable spend · 100 cells
A sample week. Avoidable is an upper bound: each event priced at the context it carried, capped at what the session cost. See yours with ellie insights me.
Review
4 in 5 review comments are never acted on, across 225,000+ in the Token Maxxing Report. Ellie reviews the working copy before you push and hands each finding to your agent.
$ ellie review · 3 files changed
findings3 found · 0:41
Staged and unstaged changes by default. A clean working copy falls back to the branch against its base.
CRITICALattribution.ts:74
+ return new Date(Date.UTC( + d.getUTCFullYear(), …));
The evidence, the code at the line, and the suggested change, checked against your team’s review guidelines.
$ ellie review --prompt-only
In src/session/attribution.ts, line 74: the day boundary is computed in local time and compared in UTC. Normalise to UTC first. Make only this change.
Claude Code applies the change1 file
With --prompt-only, each finding prints as a prompt your agent can run. In the app, c copies it.
Beyond review
The rest of Entelligence answers in the same terminal. Incident and security commands take --json, so your agent can read them too.
$ ellie incident list
ellie incident show gives the root cause and the recommended fix. Fix Now opens the pull request.
$ ellie security summary · acme/backend
82security score41 issues · 6 categories
The suggest command returns the exact code change for a finding, and issues gives the file and line your agent should edit.
$ ellie docs index
$ ellie docs get 12 --section "Retries"
Pages marked gotchas list mistakes the team already fixed. Pull them all to disk with ellie docs pull.
Review memory
When Claude Code, Codex or Cursor reads a file with review history, Ellie adds the earlier finding to the agent’s next turn. Nothing blocks a tool.
ellie hooks review-guidance, and off with ellie hooks remove review-guidance.claude code · sample session
>Fix the daily rollup. Sessions near midnight land on the wrong day.
● Read src/session/attribution.ts188 lines
Past review finding · PR #2606 · line 741 of 3
Day boundary computed in local time, compared in UTC. startOfDay() uses the runtime zone; started_at is UTC. Build the boundary with Date.UTC.
ellie rule feedback adv-2291 useful --acted-on yes
●An earlier review caught this same bug in PR #2606. I’ll build the boundary in UTC, as it suggests.
● Update src/session/attribution.ts+2 -1
Integrations
Ellie finds the coding agents on your machine and writes its MCP server, hooks and skills into each one’s own config. Running ellie hooks remove takes any of it back out.
GitHub CopilotModel routing, and sessions read for insightsnoneThe entelligence-context server answers from the Entelligence World Model. Install it with ellie mcp install.
claude code · sample session
>Add idempotency keys to the charge retry path.
● Read billing/charge.ts
The agent asks for the team’s context on the files it is about to change, then reads what that context points to.
Install
The installer checks the published sha256 before it installs anything, never asks for sudo, and writes only under $HOME.
ellie update --rollback restores the previous binary.1Install
$ curl -fsSL https:/$ irm https:/2Sign inthe only step that opens a browser
$ ellie auth login3Wire your agents
$ ellie mcp install4Review your working copy
$ ellie reviewEllie works next to the agent you already use: Claude Code, Codex, Cursor, Kiro, GitHub Copilot or OpenCode. It reviews the working copy, reads your sessions for spend and waste, and wires routing and team context into the agents that support them.
Yes. Run ellie auth login once; it is the only step that opens a browser. Review, insights and incident data come from your organisation’s Entelligence workspace.
From your own sessions, sorted into five classes. The figure is an upper bound, not a bill: each avoidable event is priced at the context it carried, and the total is capped at what the session really cost. Read it as a ranking of what to fix first.
Yes. Incident and security commands print JSON, ellie review --prompt-only prints each finding as a prompt, and ellie mcp install gives Claude Code, Cursor and Kiro an MCP server for team context.
Only what a command you ran or a hook you installed sends: the diff under review, computed session data on sync, review memory events, and an hourly heartbeat. Raw transcripts stay local unless your organisation turns transcript upload on.
Run ellie hooks remove with the hook’s name, or ellie hooks status to see what is installed. Removal holds across updates and sign-ins.
Install Ellie, run ellie auth login, then ellie review in any repo with changes.