Running offline¶

A good dev loop doesn't depend on the network. Kampong Agents can run an agent with zero network calls: a local model for the thinking, and recorded fixtures for the tool calls.

Mock and record tool calls¶

Real HTTP calls are slow and change their answers. The --tools flag on kampong run controls how tool calls are handled:

Mode	Behavior
`live` (default)	Every tool call hits the real network.
`record`	Calls live once and saves the response to a fixture file.
`replay`	Never touches the network, and reads the saved fixture. A cache miss is an execution failure, never a silent live call.

Record once:

kampong run agent.yaml --input "order 1234" --tools record

Then replay forever, deterministically and offline:

kampong run agent.yaml --input "order 1234" --tools replay

Fixtures land in <spec dir>/.kampong/fixtures by default. Point elsewhere with --fixtures <dir>. Because they're deterministic, they make tool behavior safe to commit and repeat in tests.

Secrets never reach a fixture

Any resolved BYOK key in play is scrubbed before a fixture is written, so recording a call can't leak your key into a file you might commit.

Use a local model¶

Pair replay with a local Ollama model and the run needs no network at all. Set the spec's model to Ollama, with no API key:

model:
  provider: ollama
  name: llama3.2
  base_url: http://localhost:11434

Start Ollama and pull the model once:

ollama serve
ollama pull llama3.2

Fully offline, end to end¶

A local model plus replayed tools gives you a run that touches nothing external:

kampong run agent.yaml --input "order 1234" --tools replay

You can verify the claim. Cut your network (airplane mode, or a firewall rule) and run it again. It completes with the same output.

No silent fallback, ever

If Ollama isn't running, the run fails with a clear error. It will not quietly call a paid cloud API instead. If a replay fixture is missing, that's an execution failure, and it will not quietly make a live call. Offline means offline, and a broken assumption is surfaced, not hidden.

No telemetry¶

Nothing about a local run phones home. There's no telemetry by default, and the tool doesn't report your usage anywhere unless you explicitly opt in.

Recap¶

--tools live|record|replay controls tool calls; replay never touches the network, and a cache miss is a failure, not a silent live call.
Pair replay with a local Ollama model for a run that touches nothing external.
Secrets are scrubbed before a fixture is written, and nothing phones home by default.
A missing Ollama server or a missing fixture is a loud error, never a silent fallback.

Next: edit specs from your own editor with the canvas open, in Editing specs in your IDE.