Agentic Mode via Local API & Plugin System

[eozalp]

Expose a localhost HTTP/WebSocket API in PocketPal so third-party plugins or apps can build autonomous or semi-autonomous “agentic” behaviors without bloating the core app.

Core concept:

PocketPal runs a local API server (127.0.0.1:PORT), with endpoints for:

/v1/session — start chat sessions.

/v1/chat — send/receive messages, with optional tool schema.

/v1/chat/tool_result — feed plugin outputs back to the model.

/v1/tools/execute — optional in-app tool execution (calendar, HTTP, etc.).

/v1/models — list available models.

WS /v1/stream — token streaming + tool_call events.

Plugins can be:

Local apps/scripts (Tasker, Termux, Node.js, Python).

Remote services accessed via WebSocket or HTTP.

They implement the agent loop, call the API, and handle side effects.

Tools: Model can output structured tool_call JSON (name + args).

Either PocketPal executes safe built-in tools, or

Emits an event for a plugin to handle, then posts the result back.

Security & Safety:

Localhost bind by default; API key for access.

Consent gate per tool + per plugin.

Block SSRF (no 127.0.0.1 / RFC1918 in http_request).

Quotas, logs, kill switch for agents.

Benefits:

Keeps PocketPal lightweight — agent logic & new features live in plugins.

Faster iteration — plugins can be updated independently.

Works cross-platform — same API usable on Android, desktop, or server builds.

Example Flow:

1. Plugin creates a session: POST /v1/session.

2. Sends a prompt with tool schema to /v1/chat.

3. Model responds with tool_call.

4. Plugin executes the action, posts result to /v1/chat/tool_result.

5. Repeat until finished.

Rollout Plan:

1. Phase 1: /v1/session, /v1/chat, WS /v1/stream.

2. Phase 2: /v1/tools/execute + safe http_request handler.

3. Phase 3: Consent UI, quotas, logs, plugin manager page.

4. Phase 4: Optional OpenAI-compatible routes for SDK compatibility.

[ This post copy pasted from a human-ai conversation as a summary ]

[eozalp]

Mnn app exposing an API to local host. But it's not compatible with openai specification.

For adventurous souls there is llxprt app in GitHub. A fork of Gemini cli. It has tools and tool calls. If you have right setup a LLM exposed an API and llxprt you could have an ai development in your pocket.