# postman2api

**Repository Path**: mefaso/postman2api

## Basic Information

- **Project Name**: postman2api
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: my
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-11
- **Last Updated**: 2026-08-18

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

> ⚠️ **EDUCATIONAL PURPOSE ONLY** — This project is provided **as-is** for research and educational purposes. The author is **not responsible** for any misuse, ToS violations, or consequences resulting from the use of this software. Use at your own risk.

# postman2api

Standalone Postman AI proxy — converts Postman's agent chat API into an OpenAI + Anthropic-compatible endpoint with multi-account pooling, round-robin load balancing, and Camoufox browser-automated login.

## Quick Start

```bash
bun install
cd dashboard && bun install && bun run build && cd ..
cp .env.example .env
bun src/db/migrate.ts
bun start
```

- **Dashboard**: http://localhost:1930
- **API key**: set via `API_KEY` in `.env`
- **OpenAI**: `http://localhost:1930/v1/chat/completions`
- **Anthropic**: `http://localhost:1930/v1/messages`

## Login

```bash
python3 -m venv scripts/auth/.venv
source scripts/auth/.venv/bin/activate
pip install -r scripts/auth/requirements.txt

python3 scripts/auth/postman_login.py --email you@gmail.com --password pass
```

Or via dashboard: **Logs** panel → enter credentials → click **Login**.

## API Usage

```bash
# OpenAI
curl http://localhost:1930/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-sonnet-4-5","messages":[{"role":"user","content":"Hello!"}],"stream":true}'

# Anthropic
curl http://localhost:1930/v1/messages \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-sonnet-4-20250514","max_tokens":1024,"messages":[{"role":"user","content":"Hello"}]}'
```

## Models

`claude-opus-4-8`, `claude-opus-4-7`, `claude-opus-4-6`, `claude-opus-4-5`, `claude-sonnet-4-6`, `claude-sonnet-4-5`, `claude-haiku-4-5`, `gpt-5.5`, `gpt-5.4`, `gpt-5.2`, `auto`

Antropik `/v1/messages` accepts official Claude model IDs and normalizes them automatically.

## Long-context strategy (`SEEDING_MODE`)

Upstream Postman embeds first-turn history in a single `seedingMessages` field. That field has two hard constraints (verified by probing):

1. It only accepts **exactly one `(user, assistant)` pair** — multiple entries, even when the total stays under the limit, are rejected with a `Forbidden` response.
2. Its **total length is capped at ~10000 chars** (measured: 10010 passes / 10060 rejected).

So a first-turn history longer than ~9500 chars cannot be sent in one request. `SEEDING_MODE` controls how it is handled:

| Mode | Behavior | Cost | Trade-off |
|------|----------|------|-----------|
| `multiturn` (default) | Splits the history into ~9500-char chunks and feeds each as a separate upstream round (greedily packing whole messages, never cutting a message mid-flight), then sends the real question on the same conversation. | `N+1` quota calls | Keeps the **entire** history visible to the model. Higher latency for huge contexts. |
| `truncate` | Keeps only the tail (~most recent 9500 chars) in a single call, and prepends a note that earlier history was omitted. | 1 call | Drops the oldest context. |

Set it in `.env`:

```
SEEDING_MODE=multiturn   # or truncate
```

Additional behavior notes:

- **Chunking rule** (`multiturn`): messages are packed greedily per chunk — a chunk is closed only when the next whole message would exceed the cap, so messages are never split. When a single message itself is larger than the cap (rare), only that message is truncated and annotated.
- **Retry**: each feed round retries on transient failures (network / timeout / 5xx), up to 2 retries with exponential backoff (500ms, 1s). Auth (401/403), quota, and business rejections (`Forbidden`, consent) are **not** retried. Consecutive transient failures on multiple rounds abort feeding and fall back to truncation.
- **Fallback**: if the very first feed round fails, feeding is abandoned and the request falls back to `truncate`-style behavior (tail kept + note) so the request doesn't die.
- **Boundary**: when the context fits within ~9500 chars, `multiturn` sends it as a normal single request (no extra calls), identical to `truncate` except without the omission note.

## Features

- OpenAI `/v1/chat/completions` + Anthropic `/v1/messages` protocol
- SSE streaming with thinking/reasoning tokens
- Multi-account pool with round-robin
- Auto-switch on quota exhaustion
- WebSocket real-time dashboard
- Camoufox browser-automated Google OAuth login
- Postman signup onboarding automation (fill form, start trial)
- SQLite request logging

## Architecture

```
Client → Hono API → Account Pool (round-robin) → Postman Provider → Postman API
                         ↕
                    Dashboard (React) ← WebSocket
```

Bun + TypeScript + Hono + Drizzle/SQLite + React/Vite. Python + Camoufox for browser auth.

---

> ⚠️ **EDUCATIONAL PURPOSE ONLY** — This project is provided **as-is** for research and educational purposes. The author is **not responsible** for any misuse, ToS violations, or consequences resulting from the use of this software. Use at your own risk.