S
soulboy06/forktape
Deterministic time-travel debugger, dual-protocol (LLM SSE + MCP) flight recorder & LangGraph counterfactual…
View on GitHub ↗Link checked 2026-11-14 · HTTP 200 · opens the vendor's own site
Install & config
Client config (built from the install command this repo gives)
{
"mcpServers": {
"soulboy06-forktape": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"]
}
}
}Parsed from the README we fetched on 2026-11-13; we have not executed it.
command/args come from the install command the repository states. When it is ambiguous, the project README wins.
VS Code (.vscode/mcp.json - the key is servers)
{
"servers": {
"soulboy06-forktape": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"]
}
}
}Same server, but VS Code keeps it under <code>servers</code>.
Claude Code (command line)
claude mcp add --transport stdio soulboy06-forktape -- npx -y @modelcontextprotocol/server-github
Claude Code does not read a config file; add the server with this command.
Key names differ by client: Clients & CLI
- Star count
- 0 ★ · soulboy06/forktape · read from the GitHub search API · (checked 2026-11-13 20:00:50)
- Source
- read directly from its GitHub page · type: open-source project
- Harvest history
- first seen 2026-11-14 · last seen 2026-11-14 · seen 1× · listed by 2 sources
- Score
- score 203 · 2 sources · repo
- Install
npx -y @modelcontextprotocol/server-github
FAQ (answers come from our records)
How do I install it?
The repository provides this install command:
npx -y @modelcontextprotocol/server-github. It comes from the repository itself; we cannot guarantee it runs as-is in your client.Who is it for?
We filed it under these tags: 开源项目. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
repository
soulboy06/forktape；first catalogued 2026-11-14；last seen 2026-11-14。There is no official certification or quality rating here — we record sources and dates only.# 📼 ForkTape (ftk)
以下摘录自该仓库的 README（片段） · 在 GitHub 读完整 README →
Deterministic Time-Travel Debugger, Dual-Protocol (LLM SSE + MCP) Flight Recorder, and LangGraph Counterfactual Root-Cause Doctor for AI Agents.
English · 简体中文 · ForkBench-50 Benchmark · Contributing
> Trace viewers (LangSmith, Langfuse, Phoenix) show you what crashed at Step 14.
> ForkTape rewinds time to Step 5 in
## ⚡ Try the Interactive Time-Travel Studio in 10 Seconds (Zero API Keys)
ForkTape ships with an embedded React Web Studio and 4 realistic multi-step production failure tapes (42-tool MCP schema overload, 18KB Lost-in-the-Middle context rot, silent empty-RAG hallucination, and dynamic timestamp Prefix KV Cache thrashing):
bash
# Run instantly from source with uv:
git clone https://github.com/soulboy06/forktape.git
cd forktape
uv sync --extra dev
uv run forktape demo
Or once installed via pip / uv tool:
bash
pip install forktape
forktape demo
Open the printed localhost URL (http://127.0.0.1:18787) to inspect token X-rays, run the LangGraph Biopsy Doctor, fork any step with BM25 tool pruning or context compaction, compare trajectories side-by-side, and export runnable pytest suites—100% offline with zero API keys required.
## 🤔 Why ForkTape?
When a 15-step agent fails at Step 14, the root cause almost never happened at Step 14. It happened upstream—at Step 5 when a noisy tool dumped 45KB of unpruned JSON into context (Lost-in-the-Middle), or when 64 overlapping MCP tool schemas overwhelmed tool selection, or when an empty vector search caused a downstream ID hallucination.
Re-running Steps 1..13 from scratch to test every prompt or context hypothesis is:
1. Slow & Expensive: Waiting 90 seconds and burning 40k+ tokens per edit.
2. Non-Deterministic: Sampling jitter changes Step 3, invalidating your Step 5 A/B test.
3. Dangerous: Re-executing live MCP tools (stripe_refund, delete_row, send_slack_alert) triggers duplicate real-world side effects.
| Capability | Cloud Trace Viewers (LangSmith / Langfuse) | HTTP Cassettes (VCR.py) | ForkTape (ftk) |
| Zero-SDK Streaming SSE Interception | Requires SDK / OTEL wrappers | Breaks on chunked LLM SSE | Lossless OpenAI & Anthropic SSE Assembly |
| MCP (stdio + HTTP) Flight Recording | Partial / Manual spans | No stdio JSON-RPC support | Native stdio & HTTP Bridge + Dry-Run Write Sandbox |
| Time-Travel Fork @ Step K | Single-prompt playground only | Static replay only |
1. Dual-Protocol Zero-SDK Flight Recorder (forktape.interceptor)
Point any agent (LangGraph, OpenAI Agents SDK, AutoGen, CrewAI, Claude, custom loops) at ForkTape via OPENAI_BASE_URL / ANTHROPIC_BASE_URL or forktape record -- python agent.py.
Wraps local stdio MCP servers (forktape mcp-wrap -- npx ...) and HTTP MCP endpoints, recording JSON-RPC tools/list and tools/call spans while blocking destructive mutations (delete_, drop_, send_, refund_) during counterfactual dry-runs.
Automatically scrubs API keys/Bearer tokens before persisting to local SQLite WAL (.forktape/tapes.db).
2. Canonical Fingerprinting &
| Evaluation Metric | Naive Tail-Log Baseline | ForkTape (LangGraph Biopsy Doctor) |
| Exact Causal Drift-Step Localization | 0.0% | 100.0% |
| Root-Cause Category Accuracy | 22.0% | 100.0% |
English · 简体中文 · ForkBench-50 Benchmark · Contributing
> Trace viewers (LangSmith, Langfuse, Phoenix) show you what crashed at Step 14.
> ForkTape rewinds time to Step 5 in
## ⚡ Try the Interactive Time-Travel Studio in 10 Seconds (Zero API Keys)
ForkTape ships with an embedded React Web Studio and 4 realistic multi-step production failure tapes (42-tool MCP schema overload, 18KB Lost-in-the-Middle context rot, silent empty-RAG hallucination, and dynamic timestamp Prefix KV Cache thrashing):
bash
# Run instantly from source with uv:
git clone https://github.com/soulboy06/forktape.git
cd forktape
uv sync --extra dev
uv run forktape demo
Or once installed via pip / uv tool:
bash
pip install forktape
forktape demo
Open the printed localhost URL (http://127.0.0.1:18787) to inspect token X-rays, run the LangGraph Biopsy Doctor, fork any step with BM25 tool pruning or context compaction, compare trajectories side-by-side, and export runnable pytest suites—100% offline with zero API keys required.
## 🤔 Why ForkTape?
When a 15-step agent fails at Step 14, the root cause almost never happened at Step 14. It happened upstream—at Step 5 when a noisy tool dumped 45KB of unpruned JSON into context (Lost-in-the-Middle), or when 64 overlapping MCP tool schemas overwhelmed tool selection, or when an empty vector search caused a downstream ID hallucination.
Re-running Steps 1..13 from scratch to test every prompt or context hypothesis is:
1. Slow & Expensive: Waiting 90 seconds and burning 40k+ tokens per edit.
2. Non-Deterministic: Sampling jitter changes Step 3, invalidating your Step 5 A/B test.
3. Dangerous: Re-executing live MCP tools (stripe_refund, delete_row, send_slack_alert) triggers duplicate real-world side effects.
| Capability | Cloud Trace Viewers (LangSmith / Langfuse) | HTTP Cassettes (VCR.py) | ForkTape (ftk) |
| Zero-SDK Streaming SSE Interception | Requires SDK / OTEL wrappers | Breaks on chunked LLM SSE | Lossless OpenAI & Anthropic SSE Assembly |
| MCP (stdio + HTTP) Flight Recording | Partial / Manual spans | No stdio JSON-RPC support | Native stdio & HTTP Bridge + Dry-Run Write Sandbox |
| Time-Travel Fork @ Step K | Single-prompt playground only | Static replay only |
1. Dual-Protocol Zero-SDK Flight Recorder (forktape.interceptor)
Point any agent (LangGraph, OpenAI Agents SDK, AutoGen, CrewAI, Claude, custom loops) at ForkTape via OPENAI_BASE_URL / ANTHROPIC_BASE_URL or forktape record -- python agent.py.
Wraps local stdio MCP servers (forktape mcp-wrap -- npx ...) and HTTP MCP endpoints, recording JSON-RPC tools/list and tools/call spans while blocking destructive mutations (delete_, drop_, send_, refund_) during counterfactual dry-runs.
Automatically scrubs API keys/Bearer tokens before persisting to local SQLite WAL (.forktape/tapes.db).
2. Canonical Fingerprinting &
| Evaluation Metric | Naive Tail-Log Baseline | ForkTape (LangGraph Biopsy Doctor) |
| Exact Causal Drift-Step Localization | 0.0% | 100.0% |
| Root-Cause Category Accuracy | 22.0% | 100.0% |
Listed badge (optional)
Copy this line into the README if you want the listing mark. Optional, and it does not affect being listed.
[](https://anclum.com/en/s/soulboy06-forktape/)
- Source
- github · https://github.com/soulboy06/forktape
- Browse by field
- Open source
- Listed by
- 2 independent sources
- added
- today
More like this
Filed under
Every tag opens the full list for that filter. Counts are the live catalogue figures.