strettoRead ahead of your agent.

stretto learns, from your agent's recorded tool calls, which reads come next and where their arguments come from. An MCP proxy then makes those reads for it, in the same tool result.

Watch the explainer

Get started

Why stretto?

How stretto reads ahead, learns a flow and decides a lookup, with the live results and their scope. Or step through it at your own pace.

Deep diveThe math, and the probabilistic programsThe event stretto estimates, the rule that acts on it, the counts and alpha’s posterior in fugue, and how the estimate held up in replay and live.The paper →

Deep diveThe consoleServers, sessions, flows and jobs, driven on the real console: a connection test, a session’s decisions, a threshold moved, an audit run from the page.The console guide →

Any MCP server, any host

stretto-proxy takes the server's place in your host's configuration, and runs it as a command or reaches it over Streamable HTTP. It forwards every message and records each session.

Integrations

Reads only, so the worst case is a detour

A flow calls only tools it learned as lookups and that the server does not mark as writes. A wrong guess costs one extra lookup, never an action.

Lookups and detours

No model of its own

The reach decider scores each lookup by its chance of use before the agent's next write, counted from traces. It needs no key. Live on 28 τ²-bench tasks, it cut Claude Sonnet 5's LLM turns by 20.5%, Claude Haiku 4.5's by 22.4% and GLM-5.3's by 27.9%.

Deciders

A flow is a file you review

flow-show renders a flow for a reviewer. flow-diff lists what changed and exits with 1 when a change needs review, so a flow goes through a pull request like code.

Audit and review

Shadow first, then promote

In shadow mode a flow decides and logs, but looks nothing up. stretto promote then keeps it to the sites where its lookups were the agent's own.

Shadow mode

Compile once where no user speaks

stretto-procedure runs a compiled workflow, writes included, with no model, and hands back what its outcome check cannot confirm. On τ²-bench's solo telecom tasks, 39 of 40 held-out tasks passed with a model on its 4 hand-backs.

Procedures

How it works

stretto sits where MCP already puts a server. It records what your agent does, learns which reads follow which calls, and serves those reads behind the agent's own calls.

What the agent sees

The agent searches a folder of notes for the October meetings. The flow reads both meetings behind that one call, and the agent gets its own result with one more text item:

text

--- Also looked up automatically (current results; no need to repeat these calls) ---

read_text_file {"path":"/home/me/notes/meetings/2026-10-06.md"}:
# 2026-10-06
Gamma design review moved to 10-09.


read_text_file {"path":"/home/me/notes/meetings/2026-10-13.md"}:
# 2026-10-13
Delta owner: Priya.

The agent made 1 call where it had made 3. The walkthrough runs this loop end to end on the official MCP filesystem server, with no key.

See it in the console

The console is a web app on your machine, over the files stretto writes. It shows each session as a timeline, with the reads stretto made after each call and why, and each flow as a graph you can try at another threshold. It also has your servers with their host configuration, and the CLI's jobs, run from the page.

A served session in the console: the customer's request, the agent's first call, and the three lookups stretto read ahead after it, each with its probability against the threshold

The console's guide shows how to run it, from a checkout or in a container.

The evidence

Every number comes with the kind of evidence behind it, and the script that recomputes it.

20.5%fewer LLM turns for Claude Sonnet 5 on 28 τ²-bench retail and airline tasks (16.5–24.4%), live and pre-registered, three trials; 22.4% for Claude Haiku 4.5 and 27.9% for GLM-5.3 10.1%fewer LLM turns for GLM-5.3 and Claude Haiku 4.5 in AgentDojo's own environment, in the suites with reads to take (5.8–14.0%), live, with passes unchanged 86.4%of retail's read-only ceiling taken across nine agents it never saw, 10.2 points more than a next-step speculator, in replay 39 of 40held-out solo telecom tasks passed by a compiled procedure with a model on its 4 hand-backs, at 0.48 LLM turns per ticket against 15.9

Read why stretto for where these come from and what they do not show, or the paper (PDF) for the model behind them.