PasClaw

memory & reranking setup

Local semantic search + optional cross-encoder reranking. Runs on-device; nothing leaves the host. Model downloads happen here.

Semantic memory search (local embeddings)

Reranking (2nd-stage, sharper ordering)

distilled facts export .md

workspace/memory/

Pick a file to view, or search memory.

…

Pick a file to view.

Name	Command	Args	Status

Id	Spec	Skill	Args	Channel	Status

Search the pasclaw.dev Code Vault for Object Pascal samples and components.

disconnected

Sessions

—

Turns

—

Tool calls

—

Input tokens

—

Output tokens

—

Cache read tokens

—

Cache created tokens

—

Truncation bytes saved

—

Tokens by provider

Provider	Tokens

Tokens by model

Model	Tokens

Loading…

Workers connect outbound to /v1/relay/poll, run inference locally, and POST results back — WebGPU / phone / shared home-GPU inference with no inbound ports.

How it works

Workers advertise capabilities via X-Relay-Capabilities, drain pending requests off this gateway's queue, and POST results to /v1/relay/respond/<id>. See docs/providers-relay.md for the wire protocol.

Connected workers

—

Pending

—

In-flight

—

Completed

—

Failed

—

Gateway URL Defaults to this page's origin. Override if workers reach the gateway through a different URL (tunnel, proxy, public hostname). Relay worker token Per-process credential that unlocks only /v1/relay/*. Regenerated each pasclaw serve/gateway startup. With "show" off, snippets use the placeholder <PASCLAW_RELAY_TOKEN> so you can screen-share safely.

In-browser worker

Run a relay worker in this tab: Connect loads @mlc-ai/web-llm from esm.run and downloads the model's weights on first use — nothing third-party is fetched until you click.

Security & tool-call notes

The module runs in a dedicated Web Worker (no window / localStorage, so it can't read this page's stored gateway token) and talks to the gateway with a relay-scoped token that unlocks only /v1/relay/* — a compromised model can't impersonate you. Closing the tab disconnects.

Tool calls: WebLLM only accepts the tools argument for the Hermes-2-Pro / Hermes-3 entries flagged "tool-calls supported". Other models still work for text-only turns; if a turn carries tools the worker strips them and retries (text replies come back, no structured tool_calls).

Model Worker id

idle -- pick a model and click Connect

0 served

(log)

Connect a worker from another host

Connected workers

Worker id	Capabilities	Requests	Last seen

Drag a node by its title to move it; drag from a node's right ● to another node's left ● to connect. Click a node to edit its args.

Edits write config.json on the server and apply on the next restart. Secrets are shown as ••• — leave them to keep the current value, or type a new one to change it (you can set keys here, never read them). Use Raw JSON to add or remove list entries (providers, MCP servers, …) the form can't.