Drop-in capture
Change one base URL. Every request and response — streamed or not — captured byte-for-byte with timing, TTFT, and token counts. Unknown traffic passes through untouched.
See what your agents are actually doing.
vessel sits between your LLM clients and any backend — Ollama, OpenAI, Anthropic, or anything compatible. Change one base URL and every request is captured live: prompts, responses, timing, token counts. Search it, replay it across other models, score what comes back. One small binary. No account. Nothing leaves your machine.
Finding the latest release for your operating system…
Or install with a package manager
Homebrew — macOS (Apple Silicon) and Linux
brew trust spenceclark/tap && brew install spenceclark/tap/vessel
Scoop — Windows
scoop bucket add spenceclark https://github.com/spenceclark/scoop-bucket && scoop install vessel
Find the one that went wrong
Full-text search across prompts and responses. Filter by model, backend, or agent tag. Truncations, errors, cold model loads, and tool calls a model fumbled into plain text are badged, not buried.
Tag agents, name runs
Each agent sends its own tag and each run its own session — a header, a path prefix, or a LangGraph callback. Filter, compare, and report per agent, per run.
Replay, fan out, compare
Re-send a captured request to a different backend or model — or to eight at once, or one model across a sweep of temperatures — and see every response beside the original with timing and token deltas.
Score it, rank it, report it
Rate responses 1–5 as you compare; Reports ranks models and parameter sets by score and win rate, and charts tokens, tok/s, duration, cache efficiency, and warnings by model, tag, or backend. Export any filtered list to CSV or JSONL.
Ask your AI about it
A built-in read-only MCP server lets Claude Code — or any MCP client — search and inspect your captured traffic directly. “Why did my agent stall?” is now a question you can just ask.
Quickstart
OLLAMA_HOST=127.0.0.1:4550 ollama run llama3.2
# or any SDK: base_url = "http://127.0.0.1:4550/b/openai/v1"
How Vessel is different
Vessel is a developer tool in the Fiddler / Wireshark family, not a telemetry platform: it watches the wire between your client and the model, and never touches the request.
Vessel vs Langfuse (and Helicone, Phoenix, LangSmith)
- Application-level vs wire-level. They instrument your code with an SDK and show you spans; Vessel observes the HTTP and shows you bytes. No SDK, no language binding — anything with a base URL, including tools you didn’t write (Open WebUI, aider, Cline).
- Language-neutral. Langfuse is Python/TypeScript SDKs; Vessel is one base URL from any language, or a header, or a path prefix.
- Scope. They’re multi-user, hosted-or-self-hosted, with dashboards and evals for a team. Vessel is single-binary, single-user, local SQLite. If you need team dashboards and prompt management, use them; Vessel exports CSV/JSONL if you later want to feed one.
- A dev tool, not an enterprise reporting package.
Vessel vs LiteLLM (and other gateways)
- A gateway routes and transforms: load-balancing, key management, format translation, retries. Vessel deliberately forwards byte-for-byte and never mutates. Different job; they stack — put Vessel in front of LiteLLM to see what the gateway actually sent.
- If your question is “what did my agent send?”, a gateway’s logs are one more layer between you and the answer.
Vessel vs mitmproxy / Charles / Proxyman
- They’re general HTTP interceptors: you get raw JSON and streaming chunks, and you do the parsing. Vessel speaks the three wire formats natively — it reassembles the stream, counts tokens (exact when the backend reports them), measures TTFT, flags cold loads and truncations, and lets you replay and compare. Wireshark has dissectors; this is the LLM dissector.
- No certificate install, no MITM: Vessel is a reverse proxy the client is pointed at.
Vessel vs “just log it in my code”
- Works before you’ve written any code, and for code that isn’t yours.
- Sees what actually went on the wire after the SDK’s defaults, retries, and transformations — the “is it my code or the model?” question that logging inside the code can’t answer.
Frequently asked questions
Does it change my requests?
No. Byte-for-byte, with only X-Vessel-* headers stripped. injectStreamUsage is
the single opt-in exception, and it’s flagged on the row.
Where does the data go?
A SQLite file next to the config. Nothing leaves the machine; auth headers are redacted at rest; keys are never stored.
Which backends?
Ollama native, any OpenAI-compatible server (LM Studio, llama.cpp, vLLM, OpenRouter, the OpenAI API itself), and Anthropic Messages. Unknown formats pass through and are stored raw.