Agent proposes an action
submitClaude Code, an MCP client, or a production agent describes the intended action.
Policy before action
Stop AI agents before they take risky actions.
Cordum puts policy, approval, and redacted audit evidence in front of Claude Code, MCP tools, shell commands, workflows, and production changes.
Source-available core · Local Edge path available today · Managed fleet enforcement requires endpoint controls
Claude Code · pre-tool gate
Illustrative decision$ kubectl apply -f deploy/prod.yaml
classifierproduction_change
policyREQUIRE_APPROVAL
evidenceredacted + hashed
Held before execution
An authorized reviewer can approve or reject this exact action hash.
Compatible frameworks and clouds
Works alongside the frameworks and clouds you already run
Logos indicate ecosystem compatibility, not shipped plugins.
Hover or focus this section to pause the scrolling integration list.
- OpenAI
- Anthropic
- LangChain
- AWS
- Google Cloud
- Azure
- Hugging Face
- LlamaIndex
- Temporal
- CrewAI
- AutoGen
- Mistral
- Cohere
- Vertex AI
- Bedrock
- n8n
- Zapier
One adoption path
Start with Edge. Scale to Platform.
Edge is the concrete first experience. MCP governance expands the boundary. The control plane becomes the shared backend as rollout grows.
Start here · Cordum Edge
Govern coding-agent actions locally
Put policy and approval in front of Claude Code shell commands, file changes, and governed MCP calls before side effects.
- Claude Code hook and local agentd
- Approval-bound retries
- Redacted action evidence
Developer path available today
Start with EdgeScale here · Cordum Platform
Centralize governance for agent fleets
Move shared policy, approvals, workflows, and audit into the control plane when local adoption becomes a team requirement.
- Gateway, Scheduler, and Safety Kernel
- Workflow approval gates
- Customer-managed deployment
Platform backend for managed rollout
Explore the PlatformThe Problem
The dangerous moment is an action, not a prompt.
Coding agents and MCP clients can cross from suggestion into side effects in one tool call. A post-hoc log arrives too late.
A shell can become a production change
Coding agents can run kubectl, Terraform, Git, and deployment commands with the same credentials as the developer.
MCP turns tools into mutations
A useful tool call can also write files, change policy, or touch production state unless identity and approval travel with it.
Post-hoc logs are not a control
Finding a risky action after execution is evidence of what happened, not a way to prevent the damage.
The action boundary
Policy before the action. Evidence after it.
Start at the coding-agent boundary, expand across MCP calls, and centralize the same decision model for production agent jobs.
Coding-agent actions
Cordum Edge evaluates Claude Code shell and file actions at the hook boundary.
- Allow, deny, require approval, or constrain
- Bind approvals to exact action and input hashes
- Export redacted evidence instead of raw prompts
MCP tool calls
Gateway authentication, tenant scope, and the opt-in MCP policy gate sit before tools/call.
- Caller-scoped tool and resource discovery
- Approval holds for governed mutations
- Retry coordinates for resolved approvals
Production agent jobs
The platform centralizes policy and approval when local adoption grows into a shared fleet.
- Safety checks before dispatch
- Workflow approval steps
- Searchable decision and run evidence
How It Works
Before the action runs—not after.
The same decision pattern follows an action from intent through policy, approval, execution, and redacted evidence.
1
2
Cordum evaluates policy
checkIdentity, tenant, target, risk, and policy context are checked before execution.
3
Risky work pauses
approveThe exact action is held until an authorized reviewer approves or rejects it.
4
Allowed work resumes
executeOnly an allowed or correctly approved action continues to its execution path.
5
Redacted evidence remains
auditDecision context, hashes, and approval references stay available for review.
governed action
$ cordumctl edge claude -- --print "prepare the production deploy" action kubectl apply -f deploy/prod.yaml policy REQUIRE_APPROVAL execution held before side effects evidence redacted descriptor + action hash
The action stays held until policy and approval allow it to continue.
Current status
Available now, stated plainly.
No anonymous-customer totals, invented benchmarks, or maturity shortcuts. Follow the source, docs, and explicit rollout boundary.
Available from source
Core platform
Gateway, Safety Kernel, approvals, workflows, MCP server, CLI, and dashboard live in the public core repository.
Verify the surfaceDeveloper path available
Claude Code Edge
The wrapper, hook, local agentd, Gateway evaluation, approvals, and redacted evidence path are documented today.
Verify the surfaceBuilt-in surface available
MCP governance
Cordum tools, resources, transports, tenant checks, and approval-aware mutations are implemented; per-tool policy is opt-in.
Verify the surfaceCustomer-managed today
Enterprise rollout
Cordum is self-hosted. Fleet enforcement requires endpoint, identity, and deployment controls; support commitments belong in the order form.
Verify the surfaceFrom the blog
Browse all postsField notes for governing real agent systems.
Practical guides and technical analysis for teams putting policy, approvals, and audit evidence in front of agent actions.
Newsletter signup
Frequently Asked Questions
What does Cordum stop?
Cordum evaluates AI-agent actions before execution. Depending on policy, it can allow, deny, require approval, throttle, or constrain coding-agent actions, MCP calls, and platform jobs.
Where should a team start?
Start with Cordum Edge for a concrete Claude Code action boundary. Enable MCP governance as tool-call risk grows, then use the platform to centralize policy, approvals, workflows, and audit.
Is the local Edge wrapper an enterprise enforcement boundary?
No. The wrapper is a developer and demo path. Fleet enforcement also requires managed settings, endpoint controls, trusted binaries, and service or keychain credential bootstrap.
Is Cordum source-available?
Yes. The core is source-available under BUSL-1.1, while the CAP v2 protocol is Apache-2.0. Review the code and validate the operating model before rollout.
Choose the next step
Edge first. MCP next. Platform when shared control matters.
Your AI coding agent is about to run a dangerous command. Cordum stops it.
Start with the Claude Code action boundary. Expand to MCP governance. Bring the same policy and approval model into the platform when your rollout becomes shared infrastructure.
Explore the Platform