Engineering Delivery
The Claude Code place
The Claude Code place shows how your teams use Claude Code, what it delivers, whether teams keep the standards you set, and anything worth a security look. You find it under Claude Code on the app's rail. It reads Claude Code's own telemetry; Connecting Claude Code explains how to send it.
Every number is at team level. A team smaller than the smallest team shown on its own (3 engineers by default, set in Settings) is not shown on its own, so no one can be singled out. Prompt text is never read.
What the place shows
The place has four tabs. Each reads the period, comparison and team you chose at the top of the page.
| Tab | What it answers |
|---|---|
| Usage | Who uses Claude Code, how much, on what, and what it costs. |
| Outcomes | What Claude Code produced, and whether AI-assisted work ships faster without breaking more. |
| Standards | Whether teams use the skills you require, and stay inside what you allow. |
| Security | Flags from fixed rules, permission decisions, and what is new this week. |
When no telemetry is read yet, each tab says so plainly and says how to send it. Admins are offered Setup.
A tab can also say that the tables it reads are not built yet. It then says "The tables this view reads are not built yet. They arrive after the next pipeline run.", and an admin gets Run Set up again. While a pipeline run is already building them, it says "The tables this view reads are being built. They arrive when the pipeline run going now finishes."
Usage
Usage opens with a short update and four tiles, each with a weekly sparkline:
- People in a typical week: how many people used Claude Code in an average week of the period.
- Output tokens: the tokens Claude Code generated.
- Cost at API price: what the use would cost at Anthropic's API prices. It is not your bill.
- Sessions: how many Claude Code sessions ran.
Below the tiles is a chart over time. You choose three things:
- The measure, such as cost or sessions.
- The split: team, model, query source, effort, skill, subagent, plugin, MCP server, terminal, version, operating system, project, repository or language. The six largest values get their own colour, and the rest are pooled.
- The bar size: day, week or month.
A row for each value compares the last bar with the one before.
Then come the models, with the share of tokens read from cache, and what Claude Code reaches for: skills, MCP servers, subagents and tools. Claude Code in detail sums up tools, models, spend and sessions per team.
A session that runs past midnight counts on both days in the chart over time. The What happened panel counts each session once.
What happened
Choose a bar, or one colour of a bar, to open What happened for those days. It shows:
- the cost against the days just before;
- the teams;
- the tools called;
- denied calls and calls after hours;
- the sessions, biggest first, a page at a time.
Outcomes
Outcomes shows what Claude Code produced in the period: commits and pull requests by Claude Code, lines added and removed, and the share of edits accepted.
AI assisted work against the rest, by week reads both of DORA's groups together, as Throughput and instability explains. Throughput is merged changes a week. Instability is the share of changes that needed a fix.
By team gives each team a reading of its AI-assisted changes against its other changes:
| Reading | When the app shows it |
|---|---|
| Early in the J-curve | Under 20% of the team's merged changes were AI assisted, too few to read a gain from. |
| Faster, no extra fixes | AI-assisted changes merge more than 10% faster, and need a fix no more than 3 points more often. |
| Faster, more fixes | AI-assisted changes merge more than 10% faster, but need a fix more than 3 points more often. |
| Slower so far | AI-assisted changes merge more than 10% slower. |
| About the same | None of the above. |
A team needs at least 3 merged changes in the period to get a reading. There are no tiers: each team's AI work is read against its own other work. The J-curve explains why an early dip is expected.
Standards
Standards shows the policy you set in Settings at work:
- Required skills, by team: the share of each team's sessions that used the skills you require, against your goal. A team is read by its least used required skill.
- Outside what is allowed: use of models, MCP servers, skills or tools that are not on an allow list you turned on.
With no standard set, the tab says so, and an admin is taken to the policy.
Security
Security lists the week's flags, most serious first. Each flag says the rule that raised it and its numbers, and opens its events or its sessions. The flags come from fixed rules, never a model, and each team is read against its own past:
| Flag | Rule |
|---|---|
| New to the team | A tool, MCP server, skill, subagent, plugin or model the team used this week and not in the 8 weeks before. |
| Denied share | More than 20% of the team's tool calls in the week were denied, once the week has at least 20 calls. |
| Tool volume | Bash, WebFetch or WebSearch calls more than 3 times the team's median week over the last 8 weeks, and at least 20 calls. |
| After hours | The share of calls outside working hours is more than 20 points above the team's median week, once the week has at least 20 calls. |
| Cost spike | The week cost more than 2 times the team's median week over the last 8 weeks, and at least 50 dollars. |
| Off the allow list | Use of something an allow list you turned on does not allow. |
| Required skill below goal | A required skill used in a smaller share of sessions than your goal, once the week has at least 5 sessions. |
New to the team, tool volume, after hours and cost spike need at least 3 weeks of use in the 8 weeks before, so a team new to Claude Code is not flagged for everything being new. Working hours come from Settings and are read in UTC, because Claude Code reports no time zone. Flags are for the week the period ends in, which may still be in progress.
Below the flags, Permission decisions, and what is new shows who decided each tool call, under Decided by, and what appeared for the first time, under New this week.
By person
Everything above is at team level. An admin can also let environment admins open a team's Claude Code use to its people:
- It is off by default. An admin turns it on in the Claude Code policy.
- When it is on, only environment admins see the By person view. Everyone else keeps seeing teams only.
- Every opening is recorded in the audit log.
The Claude Code policy
In Engineering Delivery's Settings, the Claude Code policy section sets what your teams may use and what they must use:
- Allow lists for models, MCP servers, skills and tools. Everything is allowed until you pick the first entry of a list. Use of anything else shows under Standards and is raised as a flag.
- Required skills, for every team or for one team, with a goal: the share of sessions that should use each one.
- By person: the switch described above.
Every name is picked from what Claude Code has reported, with how often it was used, never typed. Skill and MCP server names only come through when Claude Code sends them; see Choose whether to send skill, MCP server and tool names.
Like every setting, the policy goes into your draft. Check what this changes lists each list, the required skills and the by person switch, before and after. Nothing changes until you apply.
Learn more
- Connecting Claude Code
- Throughput and instability
- AI, for AI spend, budgets, payback and the J-curve
- Claude Code monitoring guide, Anthropic's reference for what Claude Code sends