Claude Code executes autonomous terminal bash commands, inspects local files, and runs migrations. Without endpoint protection, database passwords and cloud tokens leave your workstation in prompt payloads. ReclaimLLM catches secrets locally before egress with native pre-execution hooks.
# Developer runs: "claude 'deploy-staging script fails on auth, help fix'"
cat scripts/deploy.sh && env | grep AWS)[ReclaimLLM Hook DLP] Intercepted before egress to Anthropic API:
AWS_ACCESS_KEY_ID=AKIA[REDACTED:KEY_ID]
AWS_SECRET_ACCESS_KEY=[REDACTED:API_KEY]
POSTGRES_MASTER_PASSWORD=[REDACTED:PASSWORD]
The Agentic Threat Model
Unlike web chatbots, Claude Code has full read and write access to developer file systems and shell runtimes. Standard network firewalls cannot parse agentic tool calls inside TLS streams without breaking functionality.
When Claude Code runs commands like printenv, docker compose config, or reads .env, plaintext secrets enter model context buffers immediately.
Claude Code frequently re-reads unchanged files or dumps thousand-line test outputs across iterative turns, driving up Anthropic token bills without adding value.
Security teams have no central log of what commands Claude Code executed, what code it altered, or what data was uploaded to external LLM servers.
Engineered for Claude Code
ReclaimLLM hooks execute natively inside the Claude Code runtime lifecycle. Secret patterns (API keys, private keys, JWTs, database URLs) are caught and redacted before network transit, without adding latency.
Automatically detects when Claude Code reads unchanged files or executes repeated tests. Caches previous outputs and sends concise diffs, reducing tool-result tokens by up to 40% on shell-heavy loops.
Sessions triggering credential or PII alerts are automatically quarantined from shared feeds. Raw blobs are encrypted with AES-256 using customer-managed keys.
Centralized audit trail captures paired bash inputs, stdout diffs, file modifications, and token costs across all engineering teams with HTTPS SIEM webhook export.
| Capability | Claude Code Alone | Claude Code + ReclaimLLM |
|---|---|---|
| .env and credential redaction before Anthropic egress | No (plaintext transmitted) | Yes (default-on hook DLP) |
| Shell output secret masking | No | Yes (automated redaction) |
| Context compression for tool outputs | No (full raw output sent each turn) | Yes (diff-on-change & compaction) |
| Centralized engineering audit log | No (local machine JSONL only) | Yes (org-wide audit dashboard) |
| Customer-held AES-256 recovery keys | No | Yes (encrypted at rest) |
| Private VPC self-hosting option | No | Yes (Docker & Helm charts) |
ReclaimLLM registers a native hook for Claude Code via rclm-hooks-install. When Claude Code attempts to read files matching .env, *.env, or .envrc, or runs shell commands that output high-entropy strings, ReclaimLLM intercepts the payload locally on the developer machine and substitutes credentials with typed placeholders like [REDACTED:API_KEY] before sending prompt tokens to Anthropic.
Yes. By preserving configuration variable names and data schemas while masking only literal secret values, Claude Code can still diagnose application errors, fix syntax bugs, and write migrations without ever seeing plaintext production credentials.
ReclaimLLM applies active context compression for Claude Code sessions—including diff-on-change file caching, test output compaction (filtering passing test suites), head/tail truncation of long build logs, and loop detection—saving up to 40% on tool-result tokens in repetitive workflows.
Yes. In addition to US and EU cloud regions, ReclaimLLM's capture and governance server can be deployed entirely inside your private VPC using Docker or Kubernetes Helm charts.
Install the CLI, run the hook installer, and eliminate terminal secret leaks immediately.