How stretto works

MCP host An IDE, a desktop app or an agent framework The agent: an LLM, calling tools

stretto-proxy Forwards every message, byte for byte, and records the session session logs

MCP server Any server Files, a database, your own API: its tools, as they are

→tools/call get_user_details {"user_id":"user_7"}

←result {"email":"c7@example.com","orders":["#W7a","#W7b"],…}

Where stretto sits

stretto-proxy wraps any MCP server. The MCP host runs it in place of the server's command, and it forwards every message both ways, unchanged. The agent keeps its prompt and its tools. Until you give the proxy a flow, it only forwards and records.

stretto init prints the host's configuration, for Claude Code, Claude Desktop, Cursor or VS Code. One proxy per server: the server's command after --, or a Streamable HTTP server with --upstream.

~/.stretto/logs/orders/20260927T101500.123Z-4242.jsonl

→find_user_id_by_email {"email":"c7@example.com"}
←user_7
→get_user_details {"user_id":"user_7"}
←{"orders":["#W7a","#W7b"],…}
→get_order_details {"order_id":"#W7a"}
←{"order_id":"#W7a","status":"pending",…}
→cancel_pending_order {"order_id":"#W7a",…} write

Sessions recorded

Each log holds every message the proxy forwarded, with timestamps: the calls, their results, and the conversation if the host shares it.

Record sessions

With the configuration stretto init prints, the proxy writes each session to a log in ~/.stretto/logs/, under the server's name. Use the agent as usual. The logs are all a flow learns from, so record the kinds of request the agent will see.

Ten of an agent's own sessions give 96% (retail) and 93% (airline) of what all of them do. Replay on τ²-bench, three agents, three orders each.

In the sessions

→ get_user_details {"user_id":"user_7"}
← {"orders":["#W7a","#W7b"],…}
→ get_order_details {"order_id":"#W7a"}

The agent's next read took its argument from the result before it.

In the flow

after get_user_details

look up get_order_details

order_id ← get_user_details at $.orders[*]

after get_order_details

look up the next order in the list

Counts and bindings, in a JSON file a person can read.

Learn a flow

stretto learn counts, across the sessions, which reads followed which calls, and where each read's arguments came from: here, the order ids in the user's record. The result is a flow, a file you review like code with stretto flow-show and stretto flow-diff.

A flow only calls tools the server does not mark readOnlyHint: false; --flow-tools narrows that to a list.

The tool result the agent receives

→ get_user_details {"user_id":"user_7"}
{"email":"c7@example.com","orders":["#W7a","#W7b"],…}
--- Also looked up automatically (current results; no need to repeat these calls) ---
get_order_details {"order_id":"#W7a"}:
{"order_id":"#W7a","status":"pending",…}
get_order_details {"order_id":"#W7b"}:
{"order_id":"#W7b","status":"pending",…}

Without

user_7#W7a#W7breply

With stretto

user_7

#W7a#W7b

reply

Serve it

Served, the flow runs after each of the agent's calls. It makes the lookups it expects the agent to need, with arguments bound from the results so far, and they ride in the same tool result. The agent sees them before it would ask, so it spends fewer LLM turns. There is no new tool and no change to the prompt.

Each block is an LLM turn; the outlined ones are the turns the lookups save. Run a flow in shadow first (stretto init --flow FILE --shadow): it decides and logs, but looks nothing up. Then stretto promote keeps the sites whose lookups were the agent's own.

The customer asked about #W7a

get_order_details #W7a
used: it answered the agent's next call

get_order_details #W7b
a detour: never used

It changes no state: a flow only reads, and reads leave the state as it was.

What each later LLM turn reads

What a detour costs

A detour is a lookup the agent did not use. Its result stays in the context, so every later LLM turn reads it again: that is its cost, in tokens. It cannot change the state, since a flow only reads: it reaches an episode only through what the agent reads, never through the tools.

On the live paired run, a detour carried about 2,530 input tokens over the rest of its episode, and a saved turn saved about 6,000. GLM-5.3 on τ²-bench retail and airline.

Make a lookup when P(used before the next write) ≥ δβ + δ ≈ 0.3

β: what a saved turn is worthδ: what a detour costs chance it is read next chance it is read before the next write

make it

after get_user_details: an order

0.64 0.94

after get_order_details: a product

0.08 0.38

When to make a lookup

The reach decider asks no model and needs no key. It makes a lookup when the chance that the agent uses it before its next write clears a threshold set by costs, δ / (β + δ). That chance, not the chance the read comes next, is the one that pays: the agent may reply first, and still read it before it writes.

Replay, τ²-bench retail, two agents: after a user's details, the habit puts an order next at 0.64, and the agents read one before their next write 94% of the time; after an order, a product: 0.08, and 38%. Live costs set the threshold: 2,530 / (6,000 + 2,530) ≈ 0.3.

Live 20–28%fewer LLM turns Claude Sonnet 5, Claude Haiku 4.5 and GLM-5.3 as the agent, each on the same 28 τ²-bench retail and airline tasks, with the reach decider, paired by task.

LLM turns per episode, live WithoutWith stretto

Sonnet 53 trials

9.92

7.88

−20.5% (95% CI −24.4 to −16.5)

Haiku 4.53 trials

8.81

6.83

−22.4% (95% CI −29.4 to −15.8)

GLM-5.31 trial

11.09

8.00

−27.9% (95% CI −35.9 to −19.1)

Retail carries the savings; in airline, where the agents make more calls at once, they are not established. Passed, of 84 episodes: Sonnet 5 65 without and 69 with, Haiku 4.5 58 and 62; GLM-5.3, of 28, 24 and 21. The design can see a harm of about ten points, not smaller ones.

The numbers as a table

Agent	Trials	Without	With stretto	Change (95% CI)
Claude Sonnet 5	3	9.92	7.88	−20.5% (−24.4, −16.5)
Claude Haiku 4.5	3	8.81	6.83	−22.4% (−29.4, −15.8)
GLM-5.3	1	11.09	8.00	−27.9% (−35.9, −19.1)

A prompt does not get there: asking the agent to batch its reads, in Anthropic's own words, cut 3.4–7.6% of turns. With that prompt in place, stretto cut 17.2% (Haiku 4.5) and 22.9% (Sonnet 5) more.

Live beyond τ²-bench: 10.1% fewer LLM turns (5.8–14.0%) in AgentDojo's Slack and travel suites, where the replay found reads to take; no effect in BFCL. GLM-5.3 and Claude Haiku 4.5, one run per arm.

The ceiling is the domain's: 3.5% of turns in WorkBench to 47.1% in AgentDojo's travel suite; 29.0% in τ²-bench. Record, 89 more agents.

Results, with scope

Live, each agent ran in Claude Code on τ²-bench's own prompts and tools, with the flow in stretto-proxy between the agent and the tools. The Claude models ran three trials of every task in each arm, with Claude Haiku 4.5 as the customer; GLM-5.3 ran one, as agent and customer, against its recorded baseline. The flows were learned from τ²-bench's published runs, not from these agents. Every other result is a replay.

The live results The claims ledger Try it: the walkthrough