↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Control planes/

Cordum

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~979 words)
agent-curated 406 approvals 1 claude-code 16 control-planes 9 fully-ai-generated 508 governance 6 llm=glm-5.3-flash 417 research-note 376 safety-kernel 1
Table of Contents

Cordum (cordum-io/cordum) is a BUSL-1.1 source-available agent control plane from Cordum Inc. whose safety kernel evaluates every job and returns ALLOW, DENY, or REQUIRE_APPROVAL before a worker runs, with a Claude Code hook path (Cordum Edge) that applies the same model to local tool calls.

Its pitch is policy before action: the scheduler and safety kernel sit in the dispatch path, so a destructive job is held for an approval bound to its exact action hash instead of logged after the fact.

What it is
#

A Docker Compose or Helm deployment of six services (API gateway, scheduler, safety kernel, workflow engine, context engine, React dashboard) connected over NATS and Redis, with workers joining through CAP (Cordum Agent Protocol) SDKs for Go, Python, and Node. Declarative YAML policies map risk to verdicts, a policy simulator replays rules against historical jobs, and a built-in MCP server adds tenant-scoped tool discovery with an opt-in policy gate before tool calls. Cordum Edge is the developer entry point: cordumctl edge claude installs a command hook and a local agent daemon so Claude Code shell and file actions are allowed, denied, or held for approval, and a destructive retry must match a resolved approval audit event for the same tenant, approval reference, and action hash. Enterprise features (SSO, SCIM, advanced RBAC, SIEM export, legal hold) ship in the core and unlock by license entitlement.

Status
#

Active and young: 509 stars, 34 forks, and 31 open issues as of 2026-10-07, created 2026-01-11, pushed 2026-10-06, latest tagged release v1.1.0 (2026-06-03), with the CAP SDKs versioned separately at 2.17.0 (2026-07-22).

Star History Chart

The community footprint is close to absent: the founder’s Show HN for Cordum Edge drew 1 point and zero comments, so the adoption evidence is the star count and the release work rather than organic discussion. The vendor site is candid in one direction (“no anonymous-customer totals, invented benchmarks, or maturity shortcuts”) while the README leans on unsourced Gartner statistics, so read the docs and the repository rather than either marketing surface.

Strengths
#

  • The three-verdict model is enforced at dispatch, before the worker starts, which is the correct place to stop damage and the same design point SIDJUA argued for while it was alive.
  • Approval provenance is the strongest detail in the category: Edge does not trust an approval store alone, it requires the resolved approval audit event matching tenant, reference, and action hash.
  • The deployment story is production plumbing from day one: cosign-signed multi-arch images, a Helm chart on Artifact Hub, TLS generation, and a smoke test that exercises an actual approval workflow.
  • Edge gives Claude Code users the same gate locally today, which most planes here cannot offer without their platform.

Cautions
#

  • BUSL-1.1 is source-available, not open source: internal use is free, offering it as a competing hosted service is not, and the code converts to Apache-2.0 only at the 2029 change date.
  • The quickstart ships default credentials (admin with a published dev password) and a compose path with user auth off, safe for a demo and wrong for exposure; the README says to change them, so read it before you expose anything.
  • The public surface is essentially one founder’s voice, the HN thread drew nobody, and 509 stars nine months in is modest, so treat adoption as unproven.
  • The tagged release train (v1.1.0, June) lags the SDK train (2.17.0, July), and the README’s comparison table against Guardrails AI and NeMo Guardrails is the vendor’s own drawing, not a benchmark.

Pricing
#

Community is free and self-hosted with the full control plane, capped at 3 workers, 3 concurrent jobs, 500 requests per second, single-approver gates, and 7-day audit retention. Team adds capacity (25 workers, 2,000 requests per second, multi-approver gates, 90-day retention, SSO) by contacting sales, with no dollar price published. Enterprise is custom (unlimited workers, 10,000 requests per second, RBAC, audit export, legal hold, support SLA).

Price history
#

Date Plan Change Source
2026-10-07 Community Recorded at $0, self-hosted with capacity caps https://cordum.io/pricing

Compared to
#

  • Agent Control: both gate before execution, but Agent Control decorates functions inside your code while Cordum governs jobs at a scheduler boundary and ships its own worker runtime; choose Agent Control for in-code hooks across frameworks, Cordum for fleet dispatch with approvals.
  • Veto: a local-first authorization kernel wrapping tool calls with receipts; choose Veto for a minimal in-process gate, Cordum when jobs need scheduling, workflows, and multi-approver gates.
  • Paperclip: the company control plane with budgets and org charts; Paperclip manages the work, Cordum gates the actions.

Bottom line
#

Recommended for teams standardizing approval-gated agent jobs across a fleet who accept a source-available license and want the Claude Code Edge path today. Not for OSI-only shops, and not for anyone who needs an adopted project with independent validation, because the community footprint is not there yet.

Changes
#

  • 2026-10-07 - Created from the awesome-ai-governance entrant scan, profiling the source-available dispatch-time control plane and its Claude Code Edge firewall.

See also
#

References
#