🧵 Stitcher Proxy

LLMs have amnesia. Stitcher is the cure.

A transparent proxy that gives any LLM infinite memory. Zero dependencies. One command.

Install

npm install -g stitcher-proxy

Or run without installing:

npx stitcher-proxy

Or one-liner:

curl -fsSL https://raw.githubusercontent.com/Djsand/stitcher-proxy/main/install.sh | bash

How It Works

You point your LLM client at Stitcher instead of OpenAI/Anthropic directly. Stitcher intercepts every request, stitches in the full conversation history from local storage, and forwards it upstream. Your LLM gets maximum context every time. Transparently.

┌─────────┐     POST /v1/chat/completions      ┌──────────────────┐
│  Your    │ ─────────────────────────────────▶ │  Stitcher Proxy  │
│  App     │    (only new messages)             │                  │
└─────────┘                                     │  1. Save to JSONL│
     ▲                                          │  2. Stitch history│
     │                                          │  3. Dedup        │
     │         Response (unchanged)             │  4. Token budget │
     └─────────────────────────────────────── ◀ │  5. Forward      │
                                                └────────┬─────────┘
                                                         │
                                                         ▼
                                                ┌──────────────────┐
                                                │  OpenAI/Anthropic│
                                                │  (full context)  │
                                                └──────────────────┘

Quick Start

# Setup (pick provider, set API key, configure token budget)
stitcher-proxy init

# Start the proxy
stitcher-proxy

# Auto-configure Claude Code, Codex, and all OpenAI clients
stitcher-proxy integrate all

Usage

Change your base_url. That's it.

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8081/v1",
    api_key="your-real-key",
    default_headers={"X-Stitcher-Session": "user-123"}
)

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "What did we talk about yesterday?"}]
)
# Stitcher injected the full history. The model remembers.

curl http://localhost:8081/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "X-Stitcher-Session: my-session" \
  -d '{"model": "gpt-4o", "messages": [{"role": "user", "content": "Continue where we left off."}]}'

Works With

Claude Code · Codex · Cursor · OpenClaw · LangChain · Vercel AI SDK · Ollama · vLLM · any OpenAI-compatible client

stitcher-proxy integrate all          # Configure everything
stitcher-proxy integrate claude-code  # Just Claude Code
stitcher-proxy integrate codex        # Just Codex

CLI Reference

stitcher-proxy                        Start the proxy
stitcher-proxy init                   Interactive setup wizard
stitcher-proxy start [--port N]       Start with options
stitcher-proxy status                 Config + session count
stitcher-proxy sessions               List sessions
stitcher-proxy sessions purge <name>  Delete a session
stitcher-proxy config                 Show all settings
stitcher-proxy config edit            Interactive config editor
stitcher-proxy config set <key> <val> Quick set a value
stitcher-proxy integrate [target]     Auto-configure integrations

Configuration

All settings are tunable via stitcher-proxy config edit:

| Setting | Default | Description | |---------|---------|-------------| | port | 8081 | Proxy port | | upstream_url | https://api.openai.com | Upstream LLM API | | max_tokens | 128000 | Token budget for stitched context | | dedup_threshold | 0.6 | Similarity cutoff for dedup (0-1) | | condense_threshold | 0.35 | Similarity cutoff for condensing | | chars_per_token | 4 | Token estimation ratio | | roll_size_bytes | 5242880 | File roll threshold (5MB) |

Config priority: CLI flags → env vars → ~/.stitcher/config.json → defaults

Under The Hood

Stitcher stores every message as a line in JSONL files. When context is needed:

1. Read backward through archived files (newest → oldest)
2. Deduplicate near-identical assistant messages via trigram similarity
3. Budget — stop when token limit is reached
4. Condense — replace older similar messages with placeholders
5. Reverse — restore chronological order
6. Forward — send the full context to upstream

Sessions auto-roll to numbered archives when files exceed the configured size.

License

MIT