> Your AI writes each MCP task as one typed program instead of calling tools one by one. Delta MCP checks it, applies it all or nothing, and keeps it in Activity. How it works # Your AI stops calling tools one by one. It writes the program. Delta MCP is a free app that sits between your AI apps and their MCP servers. Your AI writes each task as one short program; Delta MCP checks it, makes the fewest calls, all or nothing, reads the result back and keeps it in Activity. Updated 8 October 2026 · 8 minutes to read A task, as your AI writes it · from Delta MCP’s tests ``` // delta { code }: one call for the whole task const t = await tickets.ticket.get("T-1"); t.status = "closed"; t.labels.push("late"); t.comments.push({ body: "Closed: fixed in 2.3" }); return t.title; // Delta MCP turns it into the fewest calls: // update_ticket, then add_comment, all or nothing ``` ## Today, your AI does three jobs at once With MCP as it is used today, the model carries out every step itself, one by one, and pays for each in tokens. ### The processor One full inference for every step: read, decide, call one tool, wait, read again. ### The bus Every piece of data passed from one step to the next goes through its context, and is paid for again at each step. ### The memory Everything must fit in its window: every tool definition, every result, the whole conversation. 55,000 tokens of tool definitions before the first word, for 58 tools on 5 servers. [Anthropic, Nov 2025](https://www.anthropic.com/engineering/advanced-tool-use) 30–50 tools: past this, the model starts choosing the wrong one. [Anthropic, tool search docs](https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool) 52.6% first-try success for the best model on 127 real MCP tasks, with 17.4 tool calls per task on average. [MCPMark, ICLR 2026](https://arxiv.org/abs/2509.24002) 128 tools: the most VS Code accepts at once. [VS Code, issue #290356](https://github.com/microsoft/vscode/issues/290356) ## A developer doesn’t click every button Coding agents are good because code has a compiler, tests and version control. Delta MCP gives your AI the same around your services: types to write against, checks before anything runs, and a history you can roll back. | Aspect | Direct MCP calls | Through Delta MCP | | --- | --- | --- | | Who decides the calls | The model, one per turn | Delta MCP, from the whole program | | What the model loads | Every tool’s definition, every time | Three tools, and typed collections for each server | | Data between steps | Goes through the model’s context | Stays in the program | | Mistakes | Found one at a time, after calls were made | All found before the first call, each with its fix | | A failure halfway | Leaves the work half done | All or nothing | | Afterwards | The model says it worked | Delta MCP reads it back and compares | | History | None | Activity: every change before and after, and Undo | | Your tools | All there | All still there: 152 of 152 on the largest server we tried | ## One door between your AI and your servers Your AI apps talk to one MCP server, Delta MCP, on your computer. Delta MCP talks to your servers exactly as they are: nothing to change, nothing to rewrite. ## How one task runs 1 ### Write Your AI writes the whole task as one short program. 2 ### Check Delta MCP checks it before the first call and returns every error at once, each with its fix. 3 ### Plan It works out the fewest calls the server needs, in the right order. 4 ### Apply It makes them all, or none. What the server can’t do is refused up front. 5 ### Read back It reads the result back, records it in Activity, and keeps what it takes to undo. Checked before the first call ``` t.labls.push("late"); ✗ no field "labls" — did you mean labels? every error at once, each with its fix; nothing was sent to the server ``` The plan · an illustration ``` d_… · 2 calls, all or nothing 1 update_ticket T-1 status=closed labels+late 2 add_comment T-1 "Closed: fixed in 2.3" Applied · read back ✓ · undo: delta_apply { undo: "d_…" } ``` ## Deep objects, as short text Some objects nest: a test has steps, a step has actions, an action has checks. Your AI reads such an object as a few lines of text and rewrites it; Delta MCP aligns the lines, keeps each one’s identity, and turns the difference into the fewest calls. Delta MCP uses this text form only when writing it back changes nothing on the server. A document · from Delta MCP’s tests ``` // read "T-1" ticket T-1: Login broken status=open // your AI rewrites it ticket T-1: Login broken status=closed labels=late comment "Closed: fixed in 2.3" Applied d… · 1 ticket(s) · 2 call(s) · read back: as written ✓ ``` ## Nothing is lost Every tool of every server stays declared and callable as it is, by name. On Cerberus, a test platform with 152 tools, all 152 are there, in a starting context about 17 times smaller (an estimate, in characters). Any original tool, by its name · from a real run, 5 Oct ``` await cerberus.cerberus_test_folder_create({ testFolder: "RT-p7-cerberus-s1", description: "Tests automationexercise.com (API + front)" }); await world.commit(); // 152 of 152 Cerberus tools, the same way ``` ## Live, when the next step depends on the last A browser can’t be planned in advance: what to click depends on what the page shows. For that, the program runs live, one action at a time. Objects it changes are still applied together at the end, and can be undone. Live mode · an illustration ``` await web.goto("https://shop.test/form"); await web.fill({ Name: "Ana" }); await web.press("Create"); return "done"; ``` ## You decide how much it does alone For each connector: change on its own (the suggestion: everything shows in Activity), ask before deleting or sending, ask before every change, or read only. Money that leaves is apart: paying, buying, refunding, transferring and trading ask you on a card by default, and your AI can never loosen that rule. Change on its own Suggested Ask before deleting or sending Ask before every change Read only ## Code execution with MCP, ready-made Anthropic and Cloudflare have both shown that agents use MCP far better by writing code than by calling tools, and both name the price: a safe place to run that code, and the checks around it. Delta MCP is that, ready-made, for every MCP server you already use, on your computer. One workflow went “from 150,000 tokens to 2,000 tokens”, a 98.7% saving, once the agent wrote code instead of calling tools. Anthropic, [Code execution with MCP](https://www.anthropic.com/engineering/code-execution-with-mcp), 4 November 2025 More than 2,500 API endpoints exposed to an agent in about 1,000 tokens, by giving it a typed SDK and asking it to write code. Cloudflare, [Code Mode](https://blog.cloudflare.com/code-mode-mcp/), 20 February 2026 ## What it doesn’t do **Save on every task.** A single one-step question costs more, because your AI reads Delta MCP’s description. The gain is on tasks with several steps. **Undo everything.** A message sent or money gone can’t be taken back, and a deleted item comes back as a new copy. Delta MCP says which, before and after. **Stop a determined AI.** It protects you from your AI’s mistakes; your AI still acts with your rights. Measured so far with Claude Code only. ## Keep reading - [How to reduce MCP token usage, and why MCP burns tokens](https://www.mcpdelta.com/guides/reduce-mcp-token-usage) How to reduce MCP token usage: measure it with /context, see the two costs (tool menu and round trips), and compare every fix, Tool Search included. - [MCP vs CLI for AI agents: where the token bill comes from](https://www.mcpdelta.com/guides/mcp-vs-cli) MCP or CLI for your AI agent? What each costs in tokens, what the measurements show, when a CLI wins, when MCP does, and where code beats both.  ## See it on your own tasks Free, two minutes to set up, and you can put everything back in one click. [Download for Mac](https://www.mcpdelta.com/download) Also for [Windows](https://www.mcpdelta.com/download#windows) [Linux](https://www.mcpdelta.com/download#linux) Free · No account · Apple chip or Intel