Chakali for defense

Operate AI with no
connection to anything.

Disconnected operation is not a deployment option here, it is the baseline. Chakali was built by a team that spent a decade securing environments that could never go online — the architecture assumes no internet, hostile inputs and compartments that must not leak.

01Why this is harder here

The constraints are not
a matter of preference.

There is no connection to lean on

A platform that degrades to a cloud call when the local path fails is not usable. Full capability has to exist with the cable pulled.

Compartments must not leak

Need-to-know is enforced between programmes, coalitions and units. AI that reasons across everything it can reach is an aggregation risk by design.

Inputs are sometimes adversarial

Documents and feeds may be crafted to manipulate whatever processes them. An instruction hidden in a captured file must not become an action.

Authority cannot be delegated to software

Consequential action requires a person who holds the authority to take it, and a record showing they did.

02What teams actually build

Real work, inside the boundary.

Your teams compose these from the Skills and Tools that installed Solutions provide. Each one runs under the same identity, policy, approval and evidence model.

01

Intelligence synthesis with provenance

Bring reporting, imagery analysis and archive material into one assessment where every claim carries its source, and confidence is stated rather than implied.

  • Source provenance
  • Compartment-scoped
  • Confidence stated
02

Alert triage and enrichment

Work the queue — classify, enrich, correlate and prioritise — so analysts spend their attention on what the automation could not resolve on its own.

  • Deterministic routing
  • Enrichment tools
  • Analyst in the loop
03

Document exploitation at volume

Process captured or acquired material in a sealed environment that leaves no trace, extracting structure and entities without the content reaching anything else.

  • Sealed execution
  • No residue
  • Entity extraction
04

Planning and course-of-action support

Several specialist Agents examine the same picture, argue the alternatives, and return a single brief with the disagreements intact for the commander to weigh.

  • Agent Teams
  • Alternatives preserved
  • Human decision
03The control model in your terms

Every consequential step
has a name against it.

  1. 01Identity and clearance

    Every request carries the person, the unit and the entitlements behind it. Access is granted to people, never to a service account acting for everyone.

  2. 02Compartment boundary

    Workspaces are sealed from one another. Retrieval cannot reach across a compartment the requester does not hold.

  3. 03Bounded tools

    Tools are typed with exact scopes, so an Agent performs the approved action and has no path to anything adjacent to it.

  4. 04Sealed execution

    Risky processing — untrusted documents, code, browsing — runs in an isolated box that leaves nothing behind when it closes.

  5. 05Command authority

    Consequential actions stop for a person who holds the authority, with the approval bound to that specific action and retained with the run.

04Where it runs

Deployment follows your risk position.

Fully air-gapped

No connectivity in either direction. Models, knowledge and execution all sit inside the enclave.

Tactical and edge

Operated on constrained hardware forward of the main environment, with the same policy and evidence model applied.

Enclave-per-programme

Separate instances per programme or coalition where the boundary between them must be physical rather than logical.

Which modelsClosed, on open-weight models inside the enclave. Chakali supports approved external providers elsewhere in the platform, and in this environment that posture is simply not offered — there is no route out of the enclave for it to use. Capability comes from the models you host and the harness around them, not from a connection. Compare the deployment patterns →

Chakali for defense

The honest question is what happens when the model is wrong and the network is gone. We are ready to answer it.