A2Agent is an API gateway. StackFerry includes these presets; users supply their own API key and choose a model available to that key.
| StackFerry client | Base URL | Protocol / authentication |
|---|---|---|
| Claude | https://api.a2agent.me |
Messages, bearer (ANTHROPIC_AUTH_TOKEN) |
| Codex | https://api.a2agent.me/v1 |
Responses, bearer |
| Grok | https://api.a2agent.me/v1 |
Responses, bearer |
| Pi | https://api.a2agent.me/v1 |
OpenAI Chat Completions, bearer |
Add a provider, select A2Agent, enter your key, fetch models, and select a model before enabling it. No default model is assumed. The Claude preset uses bearer auth; the endpoint also accepted x-api-key in the checks below.
Checked on 2026-11-02 against StackFerry base commit f847ce580e235888c7a5df48562ec387beec9d8e (with the separately proposed Claude pagination fix for the Claude discovery check). A user-supplied test key was loaded from a private local file, without storing it in source, reports, or provider configuration.
| Check | Result |
|---|---|
StackFerry listClaudeModels, bearer and x-api-key |
Both returned 22 model IDs |
StackFerry listProviderModels (Codex) |
Returned 22 model IDs |
StackFerry listGrokModels |
Returned 22 model IDs |
StackFerry listPiModels, openai-completions |
Returned 22 model IDs |
POST /v1/messages, bearer and x-api-key, anthropic-version: 2023-06-01 |
Both HTTP 200, message envelope and expected text |
POST /v1/responses, bearer |
HTTP 200, response envelope and expected text |
POST /v1/chat/completions, bearer |
HTTP 200, chat.completion envelope and expected text |
Generation checks used deepseek-v4-flash and a short text prompt. Messages used max_tokens: 2048 with thinking disabled; Responses used max_output_tokens: 2048; Chat Completions used max_tokens: 128 with thinking disabled. Discovery used StackFerry's actual provider functions with typed credentials. The discovered catalog was a single page, so it did not exercise live pagination; deterministic pagination fixtures are covered in the separate fix.
These checks cover model discovery and basic non-streaming protocol requests, not full CLI sessions, streaming, tools, routing/failover, every model, or every key's permissions. The tested key's catalog did not contain Claude or GPT model IDs; protocol support alone does not imply those models are available. Availability can change by account and over time.
This is a compatibility note, not an endorsement. Import, activation, credential storage, and routing behavior are unchanged.