Inktomi

Authority Before Commitment.

Inktomi is the first product built on Constitutional Computing. An agent may propose an action; that does not give it authority to act. Inktomi evaluates the proposal before execution, and the integrated tool verifies a single-use permit before it commits the change.

One run, every decision

1,019decisions
  • 307 permitted
  • 177 held for a person
  • 535 refused

A decision

agent-1 asked to write a file.

✕ Refused
write_file {
  "path": "demo/workspace/.env",
  "content": "[redacted: secret material]"
}

3 of 24 checks failed

  • The content is secret material.no_secret_material
  • The target file is protected.target_not_protected
  • The write needs an approval it does not have.no_approval_required

The file was never written. The decision, its policy version and all 24 checks are kept as evidence.

A replay

Same outcome.

A refused write by gemini-mcp, derived again from its recorded inputs under the rule running now.

rule 47f121a7…922b1e
From the Inktomi console, enlarged and simplified from a demonstration run.See the full console ↗
How Inktomi differs

Not a guardrail. Not a gateway. Authority checked at every step.

Most controls for AI agents either instruct the model or watch traffic as it passes. Inktomi checks each consequential action at every stage it governs, from the agent host to the tool, and the integrated tool itself refuses to act without a single-use permit for that exact action.

A call that skips the earlier checksAgentproposes1At the agent hostHost hook checks thecall before it leaves.CAN REFUSE2In transitMCP proxy sends it tothe decision service.CAN REFUSE3At the networkYour egress proxyenforces the allowlist.CAN REFUSE4At the toolVerifies the single-use permit. Last line.No permit: refused.CAN REFUSE5After the effectWorkspace watch andcredential reconciliation.DETECTS AFTERCommittedINKTOMIOne decision service behind every Inktomi checkOne governing constitution. Every decision kept as evidence.Agent proposes1At the agent hostHost hook checks thecall before it leaves.CAN REFUSE2In transitMCP proxy sends it tothe decision service.CAN REFUSE3At the networkYour egress proxyenforces the allowlist.CAN REFUSE4At the toolVerifies the single-use permit. Last line.No permit: refused.CAN REFUSE5After the effectWorkspace watch andcredential reconciliation.DETECTS AFTERINKTOMIOne decision servicebehind every Inktomi checkEvery decision kept as evidence.
Where Inktomi checks. A proposed action is checked at the agent host, in transit through the MCP proxy, and finally at the tool, which refuses any call without a single-use permit for that exact action. At the network edge, your own egress proxy also restricts destinations; it enforces and logs outside Inktomi’s record. After the effect, workspace watch and credential reconciliation detect anything the permits do not account for. One decision service sits behind every Inktomi check, so even a call that skips the earlier checks is refused at the tool. Which stages apply depends on the deployment; a coverage report shows which paths are governed.
Guardrails and model instructions

Ask the model to behave.

A model can be persuaded, confused or simply wrong, so a rule it is told to follow is only as strong as its compliance. Inktomi’s rules sit outside the model; the model never holds the authority.

Gateways and proxies

Watch the traffic that passes.

A gateway, including an MCP gateway, controls the calls routed through it. A call that goes around it, straight to the tool, is not stopped. A tool that requires an Inktomi permit refuses a direct call without one.

Policy engines and access control

Check a rule set or a credential.

They answer whether a request fits the rules, or whether a credential grants access to a system. Inktomi binds the decision to the exact action and its arguments, issues a signed single-use permit, and records which rule version decided it.

Inktomi works alongside these controls, not instead of them. Tools outside its enforcement path need their own controls; a coverage report shows which paths a deployment governs.

Not just policy

A constitution, not a rule list.

A policy engine asks whether a request satisfies a rule set. Inktomi goes further: it binds authority to the proposed action, the governing version and the evidence required before that action may become committed state.

Formally specified

Which rule wins is defined, not assumed.

Precedence between rules is specified mathematically, so the outcome does not depend on the order in which rules happen to be evaluated.

Sealed

A change is a new version, never an edit.

Rules are bound to a reviewed, signed version. Every decision records the version it was made under.

Re-derivable

An auditor checks the decision itself.

Replay derives the decision again from the recorded inputs and governing version. Agreement is shown; a changed rule or altered record cannot silently stand in for the original.

The mathematics behind Constitutional Computing

Permission for this action

Access grants capability. Authority is checked for the action.

A credential grants access to a system. It does not settle whether a particular release, payment or record change is authorized.

Inktomi evaluates the proposed call and its arguments. When the requirements are met, it issues a signed, single-use permit. The integrated tool checks that permit before acting.

1 · Propose

The agent proposes an action.

Authority is bound to the specific operation, target and arguments—not to a general ability to use the tool.

2 · Authorize

Inktomi decides.

Policy, evidence and any required human approval are bound to that exact action.

3 · Enforce

The tool verifies the permit.

A valid permit allows execution. A call without one is refused by the integrated tool.

Why the tool matters

A gateway controls traffic that passes through it. A tool that requires an Inktomi permit also refuses a direct call without one. Tools outside this enforcement path need their own controls.

Inside the console

Follow the decision. Inspect the evidence.

Console views of governed agent decisions, replay and reach.

A 78-second walk through the Inktomi console: decisions, activity, approvals, traffic and proof. No sound.

Understand why a call was allowed or refused.

Open the action, its arguments, and the integrity, identity and access checks recorded for it.

Inktomi decision record for a refused write of secret material to a .env file, showing all 24 checks and the three that failed
A refused write of secret material, with every check recorded for it. View full size ↗
Deployment

Fits around the agents and tools you run.

In the self-hosted profile, Inktomi runs on your hosts and at your network edge. Your organization keeps its policies, its identity provider and its approval responsibilities.

Your agent host

One Inktomi node

Decision service, MCP proxy, approver desk and signed journal. The agent’s own tools connect through the host hook.

Your existing infrastructure

Identity, network, monitoring

Your identity provider authenticates approvers over OIDC. Your proxy enforces network egress. OCSF events export to your SIEM.

Your integrated tools

Enforcement at execution

Each integrated tool verifies the signed permit. The evaluation establishes how every tool on the selected path performs that check.

Integrations and their status
Connects toExamplesStatus
Agent hostsClaude Code, Codex, Gemini CLI, Mistral Vibe, Hermes Agent, OpenClawAvailable for evaluation
Agent hosts, MCP onlyAntigravity: MCP tool calls are governed; its built-in tools are not yetMCP only
Tool servers (MCP)Your MCP serversAvailable for evaluation
Identity providersOIDC-based identity providers; specific products and versions are verified during evaluationProtocol support
SIEMOCSF event export; destination compatibility is verified during evaluationAvailable for evaluation
Payment, records and filing systemsYour systems of recordCustomer integration

Integration is part of the evaluation.

Start with one workflow. Run it in shadow first, recording what Inktomi would decide, then enforce. Agree tool coverage, acceptance tests, integration effort and end-to-end overhead in your environment. Inktomi’s decision engine is operational; production enforcement integrations are under development, with initial enterprise evaluations targeted for December 2026.

Coverage

Know what can stop an action. Know what reports it.

Three checkpoints can refuse in the path. Two observe and report after the fact. The deployment determines which paths are governed.

ControlWhat it doesWhen it acts
Host control Host hookChecks the host’s own tool calls before they run. It runs inside the agent’s host, so it is not the last line: an integrated tool still refuses a call without a permit.Can refuse
Tool control MCP proxy and integrated toolRoutes MCP calls through the decision service. Integrated tools verify permits.Can refuse
Network control Your proxy enforces the allowlistRestricts destinations through your egress proxy.Can refuse
Change detection Workspace watchFinds changes in the agent’s workspace.Detects after
Credential reconciliation Vendor audit logsMatches observed vendor activity against permits.Detects after

Host, MCP, workspace and reconciliation records write to the signed chain. Egress is enforced and logged by your proxy, outside that chain, so an allowlist error is not caught by the journal. Read all coverage limitations

Before you evaluate

Practical questions. Direct answers.

Does an AI model decide whether an action is allowed?

No. The decision comes from signed, versioned rules evaluated deterministically: the same inputs under the same rule version always produce the same result, and replay can derive it again. Some agent security tools use a model to judge an action’s intent. Inktomi keeps the model out of the decision.

How does Inktomi relate to endpoint security tools for AI agents?

They do different jobs and can run together. Endpoint tools find the agents running on employee devices and control what those agents do there. Inktomi binds authority to one consequential action, and the system that commits it verifies the permit before acting.

How is this different from approvals in an MCP gateway?

Many gateways grant standing approvals: an identity may use a tool, within a scope, until an expiry. Inktomi approves one exact action. The permit is bound to the operation, the tool and its arguments, lasts 60 seconds and can be used once. The next call needs its own decision.

Does Inktomi replace our existing governance tools?

No. Inktomi addresses action-level authorization, enforcement and evidence. Your identity, access, approval and monitoring controls remain in place. Its role alongside your broader governance program is established during scoping.

What does the benchmark measure?

Kihan’s recorded harness measured the governance decision at 1.22 ms median and 1.62 ms at the 95th percentile. These figures are not end-to-end workflow latency. An evaluation measures overhead on your own path.

What happens when an action is refused?

The integrated tool does not carry out the call. The decision record captures the outcome and the facts behind it. Where policy requires human approval, the action waits for an approval bound to that exact call.

What will an evaluation deliver?

A scoped workflow, agreed acceptance tests, a decision journal with replay, a coverage report and written findings. Evaluations are paid engagements. Scope and terms are agreed before installation.

See it in 30 minutes.

Explore Inktomi end to end: how it governs agent actions, handles exceptions, exposes evidence and replay, and fits your environment—with time for your questions.