Agent pushed secrets to a branch
A permissive permission policy let the agent read, stage and commit a .env file, and the commit landed before anyone reviewed the diff. Rotation and history rewrite now follow.
Secrets rotated, policy tightened
Rolling out Claude Code?
Claude Code is a capable terminal agent and a sharp knife at the same time. A single engineer running it inside a monorepo can land a feature in an afternoon, or quietly commit a secret, drop a production table, and open a pull request that nobody reviewed. We set it up with the right permissions, the right MCP servers, the right pre-commit guardrails, and the audit trail an enterprise buyer actually needs.
What it is
Claude Code is Anthropic's agentic command-line tool for software engineering, released in 2024 and now broadly used inside and outside Anthropic. It runs in a terminal, reads your repository, edits files, runs tests, drives git and shell commands, and uses Claude 4.x models to reason about what to do next. It can be extended with Model Context Protocol servers that give it access to databases, issue trackers, cloud providers and internal APIs.
It ships for individual developers on the Anthropic API, through a Claude for Enterprise plan, and against Amazon Bedrock and Google Vertex AI for teams that need their own cloud account. Permissions, approved tools, MCP servers and audit logging are configured through settings.json files at user, project and enterprise levels. The combination of agent autonomy and shell access makes guardrails more than a nice-to-have.
The honest version
No technology is all upside. This is where Claude Code earns its place and where it tends to cause trouble.
Sound familiar?
The Claude Code problems that bring people to us, and what you have once each one is fixed.
A permissive permission policy let the agent read, stage and commit a .env file, and the commit landed before anyone reviewed the diff. Rotation and history rewrite now follow.
Secrets rotated, policy tightened
A path that works on one machine is missing on another, or an environment variable needed by the server is only present in the shell after the CLI launches.
MCP boots cleanly on every machine
The account is signed up for Bedrock but not entitled for the specific Claude model, or the IAM role granting the CLI access lacks the bedrock:InvokeModel permission.
Entitlements and IAM aligned
Agent sessions running longer than they need to, engineers choosing Opus for small tasks, and no cost dashboard mean the monthly invoice doubled before anyone noticed.
Model routing, budgets and dashboards in place
A default settings.json left shell access unrestricted on every engineer's laptop, so the agent can rm, curl and push to whatever branch it pleases.
Tools scoped, approvals in place
Auto-compaction drops older messages to keep the conversation within the window, and sometimes drops the specific instruction that was holding the session on track.
Sessions split or pinned appropriately
Branch protection does not require a reviewer, the developer who prompted the agent also approves the PR, and AI-written code ships without a second pair of eyes.
Review required, humans in the loop
A dependency README or a Jira ticket contains instructions that override the system prompt, and the agent follows them unless the policy is explicit.
Input scoping and policy explicit
What we do
Call us for one failed MCP setup, or let us roll Claude Code out across the whole engineering team.
How it works
Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.
We fix the cause, not the symptom. If something is down, getting it back up comes first.
Backups, monitoring, updates and access, so the same problem can't quietly come back.
You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.
Often found together
We work on the whole stack, so the problem can be wherever it actually is.
Yes. We start with a settings.json policy that disables the obviously dangerous tools by default, wire in SSO and audit logging, configure the MCP servers your team actually needs, and run a short enablement session with each squad. Rollout is typically a week or two for a medium organisation.
First, rotate everything that was exposed, including service accounts that may have been used with those credentials. Then rewrite history to scrub the keys, force-push carefully, and run a secret scan across the whole repo. We handle the rotation and scrub, and tighten the permission policy so the same commit cannot happen again.
Almost always a path, environment variable or token scope issue. The stdio handshake is strict about JSON over a clean pipe, so anything printing to stderr during startup, or an env var that is set in your shell but not where the CLI runs, will prevent the server from registering. We walk the logs and get each server booting reliably.
If your compliance or data residency requirements point at an AWS or Google Cloud account, Bedrock or Vertex is the right choice and the CLI supports both. For teams without that constraint, the direct API is the simplest path and tends to get new features soonest. We help you pick based on your actual requirements rather than the loudest ones.
Treat any content the agent reads, including READMEs, issue descriptions, test fixtures and scraped pages, as untrusted. We configure tool permissions so the agent cannot take destructive actions without approval, scope MCP servers to least privilege, and add hooks that reject commands matching known dangerous patterns.
A written quote is free. Tell us what the team is trying to run and what has gone sideways, and a price lands in your inbox before any configuration is touched.
Free estimate
Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.
Your message is with a real person, and we'll get back to you soon.
Rather talk now? Call us: