Solution
Each laptop has different agent rules. Set them once, and see every change.
Write skills, instructions and Protect rules once, and both agents get them on every machine. On Team, one Protect policy covers every member, and History shows who changed a rule and why.
“Our blog URLs end with a dash and the tests are red. Fix it.”
Nobody typed a skill name. The skill that matches loads, and the reply names it.
“Walk me through how slugify turns a title into a slug.”
A question about the code loads no skill. At most one skill loads per turn.
You write the rules once.
Before
- A rules file in each repo
- A different file for each agent
After
- One text in Settings
- Both agents, every machine
Every request from both agents on your account carries it, on every machine. A message in the turn still wins.
An agent tries “git push origin main”.
Protected branches change through reviewed pull requests.
With Protect on, the agent reads the rule and why, and offers a branch and a pull request.
On Team, one Protect policy covers every member.
Read environment files · Block Push to protected branches · Block Change cloud resources · Monitor
Set by the org. Every member's agents follow it, on every machine.
A rule got looser. Who did it, and why?
An API key can only tighten the policy. Each entry's hash covers the one before, so an edit breaks the chain.
Six engineers, six setups. Six sets of rules.
Each agent loads skills from files on the local disk. Unless IT ships them with managed settings, every change must reach each laptop.
CLAUDE.md for one agent, AGENTS.md for the other, in each repo. One engineer's agent follows the house rules. Another's never saw them.
A rule gets looser and the agent runs a call it should not. Nobody can say who changed the rule, when or why.
Standards on the gateway. Every request carries them.
- 01Import your skills
Paste a GitHub repo URL in the console. Preview the skills, then add one or all.
- 02Write the rules once
Instructions in Settings, and Protect rules for tool calls. No file to copy into each repo.
- 03Each request gets them
The gateway adds the skill that matches and your instructions. Once you turn Protect on, it checks each tool call with the rules you pick, before the shell sees it.
Claude Code and Codex, on any laptop, send each request to the gateway.
Adds the skill that matches and your instructions. Checks tool calls with your Protect rules.
Your own Anthropic or OpenAI account gets the request and pays the model.
The rule, said out loud. Every change on record.
Blocked by your team’s policy: Push to protected branches. Protected branches change through reviewed pull requests.
The refusal the agent reads when it tries to push to main, from a real turn through the gateway. Skill choices: an offline eval on recorded turns, small sample, judged by one person. Treat it as direction, not as live accuracy. Skill eval → History docs →
Where your standards live. One place for both agents.
| Context Mode | Built into each agent | |
|---|---|---|
| Skills | One catalog per account, for both agents | SKILL.md files per agent; managed settings deploy them |
| Which skill loads | The one that matches, named in the reply | The agent decides |
| Instructions | One text per account, for both agents | Files in each repo, the home folder or managed by IT |
| Tool-call rules | One Protect policy for both agents; on Team, for every member | Per agent, through permissions and managed settings |
Where they are stronger: managed settings and plugin marketplaces reach every machine in an org, and a user cannot switch them off. Skills have no team roles yet. Method and sources →
One policy for the whole org.
On Team, one Protect policy covers every member of the org. Every member's agents follow it, on every machine.
History shows who changed a rule, how and why. Loosening changes are marked.
Put a rule in Monitor. The call runs and the log records "would block".
What this uses.
Import from GitHub once. The skill that matches loads, one per turn at most. Skills docs.
Rules you write once, in Settings, 4,096 characters at most. Every request carries them. Docs.
Related: Share skills to import your team's skill repo · Platform teams for roll-out · Shared context for facts every agent knows
Set your standards once. See every change.
npx @context-mode/cli connects your agents. 1,000 requests free, no card. Your own Anthropic or OpenAI account pays the model. We add no fee on tokens.