Collaboration and version control platform for software development and AI model hosting.
645M visits/mo-0.6%#1 Coding & Development
FreeStarting price
About vessel
vessel is a local proxy that sits between LLM clients and backend services such as Ollama, OpenAI, or Anthropic. It captures live requests and responses, including prompts, responses, timing, and token counts, without modifying the traffic. The tool supports full-text search across captured data, filtering by model, backend, or agent tag, and tagging agents and runs for organization. Captured requests can be replayed to different backends or models, including multiple targets simultaneously, with side-by-side comparison of responses, timing, and token deltas. Users can rate responses and generate reports ranking models by score and win rate, with metrics such as tokens, tokens per second, duration, cache efficiency, and warnings. Data is stored locally in a SQLite file and never leaves the machine. A built-in read-only MCP server allows integration with MCP clients like Claude Code for direct querying of captured traffic.
Key features
- Captures prompts, responses, timing, and token counts
- Full-text search and filtering by model, backend, or agent tag
- Replay requests to different backends or models
- Side-by-side comparison of responses with metrics
- Rate responses and generate ranked reports
- Tag agents and name runs for organization
- Built-in read-only MCP server for direct querying
- Exports filtered data to CSV or JSONL
Use cases
- Debugging LLM agent behavior by inspecting actual requests and responses
- Comparing model performance across different backends and parameters
- Evaluating and scoring responses for quality and efficiency
Pros
- Captures requests and responses without modifying traffic
- Supports Ollama, OpenAI-compatible, and Anthropic backends
- Local-first with no data leaving the machine
- Full-text search and filtering across prompts and responses
- Built-in MCP server for direct querying of captured traffic
Cons
- No account or cloud syncing
- Single-user only
- No team dashboards or multi-user features
- Limited to HTTP traffic observation
Frequently asked questions about vessel
Does vessel modify or alter my LLM requests?
No, vessel captures requests and responses byte-for-byte without changing the traffic. The only exceptions are the removal of X-Vessel-* headers and an opt-in flag for injectStreamUsage, which is clearly marked.
Where is the captured data stored, and does it leave my machine?
Captured data is stored locally in a SQLite file next to the configuration. Nothing leaves the machine, and authentication headers are redacted at rest while keys are never stored.
Which LLM backends does vessel support?
Vessel supports Ollama natively, any OpenAI-compatible server (such as LM Studio, llama.cpp, vLLM, OpenRouter, or the OpenAI API itself), and Anthropic Messages. Unknown formats are stored raw and passed through untouched.
How do I get started with vessel?
Change the base URL of your LLM client to point to vessel (e.g., http://127.0.0.1:4550/b/openai/v1) and run vessel as a reverse proxy. No SDK or language binding is required; it works with any tool that accepts a base URL.
Can I compare responses from different models or backends?
Yes, vessel allows replaying captured requests to different backends or models, including multiple targets simultaneously, with side-by-side comparison of responses, timing, and token deltas.
Does vessel require an account or internet connection?
No, vessel is a single-binary, local-first tool that does not require an account or an internet connection. It operates entirely offline on your machine.
