Decidealot

Decidealot

Your hardware or TypeSafe's hosted Jev. Run Laya, Von, a CLM projection head, or Jev through one TypeSafe-compatible HTTP and MCP endpoint.

At startup Decidealot downloads and verifies every enabled local bundle. It then loads only the provider selected by a request, unloads it after the configured idle period, and returns typed choice, score, and noul answers with model probabilities. CLM adds a small local projection head over one configured Qwen3-8B embeddings endpoint. It exposes the TypeSafe HTTP API and MCP Streamable HTTP from the same local container.

Contents

• Quick start
• Use hosted Jev
• Use the API
• Use MCP
• Expose MCP through a proxy
• Pick a model
• Configuration
• CLM embeddings and caching
• CUDA
• Model storage and unloading
• Agent integrations
• Docs

Quick start

You need Docker. This starts the CPU image on loopback, stores downloaded model files in one narrow host directory, and gives the container no capabilities or writable root filesystem.

model_directory="$HOME/.local/share/decidealot/models"

runtime_uid=$(id -u) runtime_gid=$(id -g) mkdir --parents "$model_directory"

docker run --detach --name decidealot --init --restart unless-stopped \ --user "$runtime_uid:$runtime_gid" \ --read-only --cap-drop ALL --security-opt no-new-privileges:true \ --pids-limit 512 --memory 8g --cpus 4 \ --tmpfs /tmp:rw,noexec,nosuid,size=128m \ --tmpfs /var/run:rw,noexec,nosuid,size=8m \ --log-driver json-file --log-opt max-size=10m --log-opt max-file=5 \ --mount type=bind,source="$model_directory",target=/models \ --publish 127.0.0.1:8080:8080 \ psyb0t/decidealot:latest

Check the service, then ask Laya to make one typed decision:

curl --fail http://127.0.0.1:8080/health

curl --fail http://127.0.0.1:8080/v1/systemone \ --header 'Content-Type: application/json' \ --data '{ "model": "laya", "state": "A proposed action would permanently delete protected data.", "questions": { "handling": { "type": "choice", "instructions": "Choose the required handling for this action.", "criteria": { "allow": "The action is reversible and does not affect protected data.", "require_review": "The action is irreversible or affects protected data." } } } }'

The first service startup downloads the enabled pinned model bundles and can take several minutes. The default configuration enables Laya and Von. Wait for /health before sending a decision. Later service starts reuse the same host directory.

An example response looks like this. The answer, probabilities, and token count vary with the model call:

{

"model": "laya", "answers": { "handling": { "type": "choice", "choice": "require_review", "confidence": 0.94, "probabilities": {"allow": 0.06, "require_review": 0.94} } }, "usage": {"input_tokens": 91, "output_tokens": 0} }

For that result, your application queues the action for human review instead of running it. Decidealot does not perform the action. Your caller checks the chosen label and applies its own probability threshold, for example allowing an action only when choice is allow and probabilities.allow is at least 0.95.

Use hosted Jev

Set DECIDEALOT_TYPESAFE_API_KEY in a private environment file or secret store. Jev enables automatically when the key is present. Call GET /v1/models and select a hosted name returned by TypeSafe, such as jev-latest. Decidealot caches the authenticated TypeSafe catalog for 60 seconds and refreshes it on the next model-listing or hosted-model request after expiry. New hosted model names need no Decidealot release. It sends the complete state and questions to TypeSafe's fixed HTTPS API, not to a local model. The key never belongs in the request body, a tracked file, or a client-side application.

For a Jev-only container, disable all local models so startup downloads no model bundles. Pass the upstream key as an environment variable from your secret store, then use the same REST or MCP endpoint shown below:

docker run --detach --name decidealot --init --restart unless-stopped \

--read-only --cap-drop ALL --security-opt no-new-privileges:true \ --tmpfs /tmp:rw,noexec,nosuid,size=128m \ --env DECIDEALOT_TYPESAFE_API_KEY \ --env DECIDEALOT_LAYA_ENABLED=false \ --env DECIDEALOT_VON_ENABLED=false \ --publish 127.0.0.1:8080:8080 \ psyb0t/decidealot:latest

The variable must already be set in the shell or supplied by your container manager. GET /v1/models confirms which Jev selectors are enabled. The API guide has a complete request and response example. TypeSafe charges for its API calls; model probabilities still need your own action thresholds.

Use the API

Send the state to judge and a bounded question. Decidealot returns typed results with probabilities.

choice questions need named criteria. score questions need an ordered criteria array whose position is the score. noul questions return a probability between zero and one. The API guide has the request rules, response shape, validation failures, aliases, authentication, and lifecycle behavior.

Put the subject in state, the question and decision rules in each question's instructions, and explicit descriptions in criteria. A label like review does not explain what should be reviewed. For noul, write the statement whose probability you want in instructions.

CLM accepts optional flat request configuration. The selected model already identifies which settings apply:

{

"model": "clm", "config": {"temperature": 0.8}, "state": "The operation permanently removes protected data.", "questions": { "review": { "type": "noul", "instructions": "This operation requires human review because it removes protected data." } } }

CLM temperature defaults to 1 and accepts finite numbers greater than 0 and at most 100. Lower values sharpen its probability distribution; they do not make the answer more correct. Laya, Von, and hosted Jev currently accept only omitted or empty config. Unsupported keys and values return 422, never silently disappear. The same config belongs in MCP tool arguments or each individual batch request. Request configuration covers the contract.

Use MCP

The same container serves MCP Streamable HTTP at http://127.0.0.1:8080/mcp. The system_one tool takes the same model, state, and questions fields as POST /v1/systemone. system_one_batch takes a requests list, each item with those same fields. list_models returns the live catalog. unload_models releases local model memory. Direct loopback clients and containers using the decidealot Docker service name work by default.

Point an MCP client at that exact URL. Its configuration format varies, but the connection values are always equivalent to this:

{

"mcpServers": { "decidealot": { "url": "http://127.0.0.1:8080/mcp", "headers": { "Authorization": "Bearer your-token-here" } } } }

Omit the Authorization header only when DECIDEALOT_API_KEY is empty. The MCP tools return structured output matching the HTTP result bodies, so an agent can inspect probabilities before it chooses the next action. A client that only supports local stdio can use the optional OpenClaw bridge described in Agent integrations. The API guide has tool inputs, output shapes, session behavior, and failure behavior.

Expose MCP through a proxy

The MCP server keeps DNS-rebinding protection enabled. A proxy, tunnel, or public DNS name must be allowed explicitly. Add the exact public Host value to DECIDEALOT_MCP_ALLOWED_HOSTS. Browser-based MCP clients must also add their exact origin, including the scheme, to DECIDEALOT_MCP_ALLOWED_ORIGINS.

For a proxy that publishes https://mcp.example.net/mcp, put these values in the Compose .env file before docker compose up -d, or pass the same variables with docker run --env:

DECIDEALOT_API_KEY=replace-with-a-real-secret

DECIDEALOT_MCP_ALLOWED_HOSTS=127.0.0.1,127.0.0.1:*,localhost,localhost:*,[::1],[::1]:*,decidealot,decidealot:*,mcp.example.net DECIDEALOT_MCP_ALLOWED_ORIGINS=http://127.0.0.1:*,http://localhost:*,http://[::1]:*,https://mcp.example.net

Keep --publish 127.0.0.1:8080:8080 when a local reverse proxy terminates TLS. The proxy forwards the request unchanged with Host: mcp.example.net. After bearer authentication, Decidealot returns 421 for an untrusted host and 403 for an untrusted browser origin. Invalid or missing bearer credentials return 401 first. Do not disable this protection or allow a broad wildcard for an internet-facing endpoint.

Pick a model

Every request must name a selector. GET /v1/models returns the same catalog at runtime.

What differs

Laya is one model family with three checkpoints. Its automati