GTP Accelerators
Working starting points for production AI agents, with human approval, audit trails and evaluation built in. Source is on GitHub.
Anthropic ClaudeMCPPython
Document-heavy workflows with Claude, humans in control
Extracts structured fields from claim documents, looks up the policy over MCP, routes each claim with auditable rules, and drafts a reply that waits for human approval.
- ✓Schema-validated extraction using a forced Claude tool call
- ✓Routing rules kept outside the model so every decision is explainable
- ✓Nothing is sent without a recorded approval; every step is audit-logged
- ✓Built-in evaluation harness for extraction and routing accuracy
LangGraphAnthropic ClaudePython
Incident response with a human approval gate
A LangGraph state graph that gathers logs, metrics and recent deploys, diagnoses the likely root cause, proposes a fix, pauses for on-call approval, then runs it from an allowlist.
- ✓Human-in-the-loop pause and resume using LangGraph interrupts
- ✓Only allowlisted actions can run, and execution defaults to dry-run
- ✓Low-confidence and informational alerts never trigger action
- ✓Swap the fixture tools for Datadog, CloudWatch, Prometheus or Kubernetes
CrewAIOpenAI or ClaudePython
Legacy code to audit-ready remediation
Three specialist agents analyse a legacy Python file, map vulnerabilities to NIST CSF, OWASP and CWE, and produce refactored code, a pytest suite, a git diff and a pull request description.
- ✓AST analysis and a 17-rule security scanner run in plain Python, no model cost
- ✓Only three model calls per run, so latency and cost stay predictable
- ✓PR description includes a human-approval checklist for security, engineering and compliance
- ✓Designed to plug into CI and extend with your own rule sets
These are reference builds demonstrated on synthetic data, not client deployments. Names of third-party products are used to describe what each build runs on and do not imply endorsement or partnership.
Want one adapted to your environment?
We tailor these to your data, systems and approval processes, and measure results on your own cases.
Talk to Us