Model providers

Model providers manages API endpoints, credentials and selected models for Claude Code, Codex, OpenCode, Pi and OMP in Paseo. It adds independent provider entries to Paseo's normal agent picker and a settings page for configuring providers and starting sessions. Existing subscription logins and agent presets remain available; custom APIs use the selected provider's API quota.

Requires Paseo 0.12.0 or newer on the daemon and clients, and each selected agent CLI already installed on the daemon's PATH. The plugin does not install external software at runtime. The interface and settings entry follow the device language, supporting English and Simplified Chinese with English fallback. Paseo 0.12.0 does not expose the app's manual language preference to plugins.

New providers can start from a DeepSeek or OpenRouter preset. OpenRouter uses Chat Completions at https://openrouter.ai/api/v1 with OpenCode, Pi and OMP; load its model catalog or enter IDs before saving. It has no hardcoded model list; missing capacities fall back to 32,768 context and 4,096 output tokens. All default settings, including the API base URL, remain editable for compatible gateways. Preset changes clear the entered key. Optional per-model mappings keep aliases in the picker while agents and connection tests send the mapped API IDs. DeepSeek, Anthropic and OpenRouter catalog metadata is retained when returned; context and output limits are configured per model row, not shared across selected models. Loading fills missing capacity fields from metadata while retaining manual values. Codex uses a native model catalog for each mapped model’s context window; OpenCode and Pi/OMP receive context and output limits; blank fields fall back to legacy provider defaults. Preset capacities are defaults, not guarantees about custom gateways.

The compact workbench supports provider search by name, endpoint, alias, API model ID or agent, plus enabled-status filters. The settings column provides list/detail navigation; a sufficiently wide content area shows the list beside its details. The page is available under Settings → Plugins → Model providers and from the Command Center, with no homepage sidebar entry. Add or edit a provider in a host popup with a summary header and separate connection, model and agent tabs. Desktop model rows use a compact table, while mobile rows keep labeled fields. Closing an edited form asks before discarding changes; pending loads and saves lock editing and dismissal. Tab changes preserve drafts, validation returns to the relevant tab, and the save footer remains visible. Choose its protocol and edit one model per form row. Load models fills the form automatically, preserving manual entries and mappings; rows can be added or removed individually, up to 100 models per provider. Anthropic Messages supports Claude Code, OpenCode, Pi and OMP; OpenAI Responses supports Codex, Pi and OMP; Chat Completions supports OpenCode, Pi and OMP. Anthropic authentication supports either x-api-key or Bearer tokens. The plugin does not translate between protocols. Changes apply to new or reopened sessions, without switching a running session's endpoint. Disabling a provider does not stop existing sessions.

Codex requires a CLI supporting codex debug models --bundled (verified with 0.160.1). Saving builds a local catalog from that installed CLI’s model instructions and tools, with the configured per-model context windows. Its native reserve remains in effect: a 1,000,000-token window currently reports 950,000 usable tokens. The output-limit field does not configure Codex or Claude Code; Claude Code also keeps its own context rules. After updating this plugin, save existing providers again and reopen sessions to refresh the generated configuration.

Credentials are stored in plaintext on the daemon under PASEO_HOME/model-providers/suppliers.json (directory mode 700, file mode 600), and in the generated Paseo provider environment configuration. Users able to read those files can access the keys. Saved keys are not returned to the interface; leaving the key blank when editing keeps the previous value. Endpoints without authentication can use an empty key. The plugin updates only its own provider aliases through the Paseo SDK and rejects collisions with other configurations. It does not read or change official authentication files.

Fetching models uses the standard /v1/models endpoint under the configured API base URL, retaining an existing version suffix such as /v2. Supply an API base URL without /chat/completions or other request paths. Responses must use a standard data array; names and optional context/output limits are retained, and Anthropic pagination is supported. Failed discovery preserves manual model entries. Requests send the provider key to the configured API endpoint and reject redirects. Errors do not expose response bodies or keys. Connection tests send a short “Reply OK.” prompt and a tool definition and may consume billable tokens; the test never executes the tool. Starting an agent sends prompts, code context and tool results to the configured provider according to that agent's behavior. The plugin has no telemetry service; the endpoint and its data policies are chosen by the user.

Model catalogs do not prove invocation permission or inference compatibility. Connection tests use non-streaming JSON, so streaming-only endpoints may fail the probe. A successful connection test checks only basic protocol response structure; streaming, end-to-end agent tool use, MCP integration and model quality are not certified. Pi/OMP models are registered through a private local extension with text input, reasoning disabled and unknown pricing; zero cost metadata means unknown rather than free. Enter context and output limits from your provider's specifications; full token counts and K/M shorthand are supported (1K = 1000 tokens). The editor uses the host modal without DOM overrides. Actual compatibility depends on the installed CLI and provider API.