toolsmelt

Formerly mcpsmith. uvx mcpsmith and servers generated by mcpsmith keep working; your saved connections in ~/.mcpsmith, MCPSMITH_* variables and keychain entries are still read. Run uvx toolsmelt setup once to swap the agent entries over.

Turn an OpenAPI file, a Postman collection, a website you log in to, an app you own or a SQL database into a few curated tools for your AI agent. New formats plug in as packages (docs/extending.md). An LLM designs the tools; you get ordinary Python code you own if you want to publish them.

uvx toolsmelt setup

What you can turn into tools

Apps and programs: only software you own or are authorized to analyze.

toolsmelt --help: the command list.

Quickstart: a logged-in website

uvx toolsmelt setup                  # once: registers toolsmelt + Chrome DevTools MCP with your agents
uvx toolsmelt browser                # opens the shared Chrome; log in to the site yourself
# restart your agent, then ask it:
#   "Record example.com in the shared Chrome and connect it as tools."
# the extract-api skill runs the rest: record, inspect, design, validate, connect

Your passwords and 2FA codes never reach the agent. Details are under Live hub.

Proven on real software

• Tusky 29.0 (open-source Android client, Retrofit): toolsmelt app-android found 113 endpoints with every HTTP method resolved (48 GET, 43 POST, 11 DELETE, 8 PUT, 3 PATCH) and no unresolved notes. Minified annotation names were resolved from the app's own request-builder class, not guessed.
• GitHub Desktop 3.6.7 (Electron): toolsmelt app-js found 33 endpoints with every method resolved. All 33 were traced to their original TypeScript line through the app's source maps.
• hn.algolia.com (public API, no login): recorded in the shared Chrome. 2 curated tools were connected through the hub and returned live results. The tools were still present after restarting the hub.
• An internal attendance dashboard (ASP.NET, cookie login): recorded while logged in. 2 tools answered the owner's questions. After logout the tools return a clear session-expired error.

Numbers are from runs on 2026-10-21 with toolsmelt 0.4.0. Reproduce the first two with toolsmelt app-android and toolsmelt app-js on those apps.

toolsmelt app-js reading an Electron app's own code.

For coding agents

No API key needed: your coding agent is the designer.

• Skills (plain Markdown, any agent can read them): skills/build-mcp-server (the whole workflow), skills/extract-api (logged-in sites into live tools), skills/design-mcp-tools (design checklist). AGENTS.md points Codex, Cursor and Gemini CLI at them.
• MCP tools (paths absolute): smith_guide, smith_inspect, smith_app_android, smith_app_js, smith_app_desktop, smith_probe, smith_draft, smith_validate, smith_generate, smith_check, smith_install, and for the hub smith_record_start, smith_record_stop, smith_connect, smith_disconnect, smith_connections, smith_login, smith_tools, smith_call.
• MCP config, any client:

  {"mcpServers": {"toolsmelt": {"command": "uvx", "args": ["toolsmelt", "mcp"]}}}
  

• Claude Code plugin (skills + MCP server): /plugin marketplace add sepehr071/toolsmelt, then /plugin install toolsmelt@toolsmelt.

MCP clients start servers without your shell's proxy variables. Behind a proxy, add "env": {"HTTPS_PROXY": "http://127.0.0.1:8080"} so smith_inspect can fetch specs.

Live hub

Skip the project: connect a plan and its tools appear in your agent right away. The toolsmelt MCP server is the hub. Connections are saved under ~/.toolsmelt (TOOLSMELT_HOME) and load again on restart. They are personal (one machine, one user); generate stays the way to publish.

From the agent: smith_inspect, edit plan.yaml, smith_validate, then smith_connect(workdir, name) and call the new tools. If your client does not show tools added after start, use smith_tools(connection) and smith_call(connection, tool, args) (read-only tools only). smith_connections lists, smith_disconnect(name) removes (CLI: toolsmelt connections).

Logged-in sites

For a dashboard or portal with no public API, the agent records the site's own calls while you use it. The extract-api skill runs the whole flow.

toolsmelt browser                                                      # shared Chrome on port 9222 (--port N, --no-proxy)
npx chrome-devtools-mcp@1.11.0 --browser-url=http://127.0.0.1:9222    # point Chrome DevTools MCP at it

1. Log in yourself in that Chrome window (passwords and 2FA never reach the agent).
2. The agent calls smith_record_start(name, hosts), browses each feature twice with different inputs, then smith_record_stop(name), which writes a HAR with cookies and tokens dropped.
3. It inspects, designs and connects the plan with auth: {type: browser, login_path: /login}. Calls reuse your Chrome session; when it expires the tool says to log in again.

Agents must not use DevTools get_network_request: it prints request headers, including cookies. generate refuses browser-auth plans because they only work with your Chrome; switch to token or API-key auth to publish.

Websites without the hub

pip install "toolsmelt[capture]" && playwright install chromium
toolsmelt capture https://example.com -o site.har --host example.com   # browse, then close the tab
toolsmelt inspect site.har --host example.com
toolsmelt draft --name example-mcp
# redesign plan.yaml, then validate, generate, check as under Export & publish

While recording, do each thing you want as a tool: search, open a detail page, paginate, apply a filter. Anything you do not click is not in the capture. --host (repeatable) keeps only the hosts you list. A DevTools "Save all as HAR" export works in place of capture. The recording browser uses HTTPS_PROXY when it is set.

A HAR holds everything the browser sent and received, including cookies and your personal data: keep it out of git and delete it after inspect. inspect scrubs what it keeps: an API key or token becomes an auth env var with its value dropped, other secret-looking headers and query params are dropped, and emails, phone numbers, JWTs and values under secret or personal field names are masked in examples and test fixtures. This is pattern-based, so read tests/fixtures/ before you publish a generated server. If a key is a public one the site sends to every visitor (a search or maps key), you may put it in the tool's fixed values.

Apps and programs

Only for software you own or are authorized to analyze.

• Desktop program: toolsmelt app-desktop prints the steps to record its traffic with your own mitmproxy (mitmdump --set hardump=capture.har); toolsmelt app-desktop capture.har --host api.example.com imports the result. toolsmelt runs nothing for you, and certificate-pinned programs are out of scope.
• Android app: toolsmelt app-android app.apk --base-url https://api.example.com reads the app's Retrofit interface. It needs a decompiler: the free jadx CLI on PATH (scoop install extras/jadx, brew install jadx), or REA's JADX jar (REA_JADX_MCP_JAR). Without --class it lists candidate classes. Before decompiling it triages the APK in about a second (zip listing and dex bytes): a Flutter or React Native app, or one without Retrofit, gets its next step right away (record its traffic, or app-js on a React Native JS bundle), and a known request-signing SDK names the sign-in provider to use.
• App that signs its own requests: plans use auth: {type: signin, provider: xiaomi} (Mi Home cloud is the first provider). The user signs in on the provider's own page through a link from the server's <prefix>_login tool (smith_login in the hub); the session goes to the OS keychain. Calling the login tool again names the account and where its data is, which catches a wrong account or region.
• Electron/JavaScript app: toolsmelt app-js <dir|app.asar|file> --base-url ... reads the calls in its shipped code.

Neither app-android nor app-js runs the app. The endpoints are unverified (no example response) until toolsmelt probe calls the GET ones live with your credentials (--writes for the rest, only after you agree). A method that cannot be read stays empty and validate refuses it.

Setup

uvx toolsmelt setup registers the toolsmelt and toolsmelt-chrome (Chrome DevTools) MCP servers, and copies the skills into each coding agent it finds (Claude Code, Codex, Cursor, Gemini CLI). Options:

• --dry-run prints the plan and writes nothing.
• --agents claude,codex limits it to those agents (default: the ones found).
• --rea also installs REA (rea-agents). It is off by default: its MCP server adds about 200 tools to each agent, and toolsmelt's app path needs only plain tools such as jadx.
• --no-skills skips the skills; --yes skips the confirmation.
• --remove undoes what setup did. Entries and skill folders you made yourself are left alone. REA stays installed (you may have had it before); the command prints the line that removes it.

Before setup first changes a config file, it copies the file to *.toolsmelt.bak. The backup is made once and not refreshed by later runs. Restart your agents afterwards. Run toolsmelt browser before using Chrome DevTools: it starts the shared Chrome the server connects to.

Claude Code plugin users already get both servers from the plugin and do not need setup.

toolsmelt install <project> registers a server made by generate with the same agents, the same way: it runs the project folder (uv run --directory), or the released package with --pypi. It touches only the entry named after the plan and refuses one of that name that runs something else unless --force. --remove takes it out again.

toolsmelt setup --dry-run prints what it would change and writes nothing.

Transforms

Plans convert formats with transform on arguments, output.transforms (date->unix, date->unix_end, unix->date, unix->datetime, minutes->hh:mm, toman->rial, rial->toman, int->str, plus optional Jalali calendar conversions) and timezone (IANA name; empty means the system's local zone).

Secrets

Credentials come from you, never the chat: toolsmelt secret set <connection> <ENV_NAME> asks with a hidden prompt and stores the value in the OS keychain (--once forgets it after one use, for one-time codes; list and delete also exist).

Export & publish

To turn a plan into a standalone server project instead of a hub connection:

uvx toolsmelt inspect https://petstore3.swagger.io/api/v3/openapi.json   # .toolsmelt/source.json + digest.md
uvx toolsmelt draft --name petstore-mcp                                  # heuristic .toolsmelt/plan.yaml, no LLM
# redesign plan.yaml (your coding agent does this, or use `design`, see below)
uvx toolsmelt validate
uvx toolsmelt generate -o petstore-mcp
cd petstore-mcp && uv run pytest
uvx toolsmelt check .                                                    # call every tool live
uvx toolsmelt install .                                                  # register it with your agents

A spec read from a local file may only have a relative server URL; validate then asks you to set an absolute base_url in the plan.

Let toolsmelt call an LLM

uvx toolsmelt design --goal "find products, compare prices, read reviews" --provider anthropic

Providers: anthropic (ANTHROPIC_API_KEY), openai (OPENAI_API_KEY), openrouter (OPENROUTER_API_KEY), ollama (no key). --model picks the model; TOOLSMELT_MODEL and TOOLSMELT_BASE_URL override the model and endpoint. The result is the same plan.yaml; review it before generate.

How it works

openapi.yaml | site.har --inspect--> source.json + digest.md
                        --draft----> plan.yaml     (heuristic starting point)
   your agent edits plan.yaml, or `design` ------> plan.yaml
                        --validate-> issues
                        --generate-> <name>/       (server project)
                        --check----> live report per tool

Source is what the API offers; Plan is the server to build; code generation reads the plan plus captured responses for test fixtures. A plan, trimmed:

name: shop-mcp
description: Search products and compare prices on Example Shop.
prefix: shop
base_url: https://example.com/api
env_prefix: SHOP_MCP
instructions: Search first, then fetch details with the id. Prices are in toman.
tools:
  - name: shop_search
    title: Search products
    description: Returns matching products with price and stock. Use shop_product for full details.
    endpoint: GET /v2/search
    args:
      - {name: query, type: string, required: true, description: "Search words, e.g. 'laptop'", to: query.q}
      - {name: max_price, type: integer, description: "Upper price limit in toman, e.g. 20000000", to: query.price_max}
      - {name: sort, description: "Result order.", map: {cheapest: price_asc, newest: date_desc}, default: cheapest, to: query.o}
    output:
      root: data.products
      fields: {id: id, title: title_fa, price: price.selling_price, in_stock: stock.available}
      max_items: 10
    example: {query: laptop}

Argument targets (to:): path.<n>, query.<n>, header.<n>, cookie.<n>, body, body.<a.b.c>. Output paths use dots, a number indexes a list (images.0.url), [] maps over a list (tags[].name). map turns friendly values the model sees into the cryptic ones the API wants.

A tool can also follow pages, or chain calls (one or the other per tool):

    # in one tool: read up to 3 pages and join the lists at output.root
    paginate: {style: page, param: query.page, max_pages: 3}   # or offset, cursor, next_url
    # in another: search, then fetch the details of the first 5 hits
    steps:
      - {endpoint: getProduct, for_each: data.products, max_items: 5, args: {path.id: $item.id}}

body_encoding (json, form or multipart) says how body.* values are sent. A plan carries schema_version: 1, unknown keys are errors (a typo like readonly: fails instead of being ignored), and toolsmelt schema prints the plan's JSON Schema for editor completion and validation.

What you get

<name>/
  pyproject.toml  server.json  README.md  LICENSE  .gitignore  toolsmelt.yaml
  src/<package>/server.py        # tools, built on toolsmelt.runtime
  tests/                         # replay captured responses; one -m live test per tool
  .github/workflows/             # ci.yml, publish.yml (PyPI + MCP Registry)
  .claude-plugin/plugin.json     # Claude Code plugin
  skills/<name>/SKILL.md         # skill for the generated server

toolsmelt.yaml is the plan, so you can regenerate. Each tool in server.py keeps its typed signature and calls toolsmelt.runtime.interpret.run_tool, the same code the hub runs.

Run a generated server (or the hub, toolsmelt mcp) over Streamable HTTP instead of stdio with --transport http --port 8000 (or MCP_TRANSPORT=http); it listens on 127.0.0.1 at /mcp. Set MCP_AUTH_TOKEN to require Authorization: Bearer <token> on every request (needed before you bind to a non-loopback --host). To publish: add a PyPI trusted publisher (workflow publish.yml, environment pypi), bump the version in pyproject.toml and server.json, tag v*.

Databases

toolsmelt inspect postgresql://reader@db.example.com/shop reads tables, columns, keys and up to 3 sample rows (personal values masked). draft writes list and get-by-key tools as plain SQL (sql: with :name placeholders, args to: sql.<name>). A read-only tool must be one SELECT (or WITH ... SELECT); this is checked by validate and again at run time, and SQLite and PostgreSQL also open the connection read-only. The password never goes into the plan: it comes from <P>_DB_PASSWORD (or toolsmelt secret set in the hub). Install the driver for your database next to toolsmelt[sql] (for example psycopg).

Tool-choice evals

toolsmelt eval --write --provider anthropic writes evals.yaml (one or two user questions per tool, plus a few that no tool should answer). Review the file and commit it. toolsmelt eval --provider anthropic then asks the model which tool and arguments answer each question, and reports tool and argument accuracy and which tools get confused. --json and --min 0.9 make it a CI gate. The tool list is the cached prompt prefix, so repeat runs are cheap.

Drift check

toolsmelt check <project> [--json] calls every tool with its example arguments and reports errors, output size in characters and output fields that came back empty. Run it nightly in CI to see when an API or site changes. It makes real calls, so tools marked read_only: false are skipped unless you pass --writes.

Runtime features

Generated servers import toolsmelt.runtime, so a fix there reaches every server on upgrade.

• Retries on network errors, timeouts, 429 and 5xx (writes only on 429, so they are never sent twice); Retry-After honoured and capped; jitter; concurrency cap.
• GET response cache, host allowlist, 5 MB body cap.
• Proxy support: <P>_PROXY takes one URL or a comma list (pool). <P>_BASE_URL overrides the base URL. System proxy env vars are ignored on purpose.
• Chrome TLS fingerprint for sites that block plain clients: pip install "toolsmelt[stealth]" and impersonate: true in the plan. Without the extra it falls back to plain httpx with one warning.
• Compact output: shape() keeps only the planned fields, caps lists (reports the total count), clips long strings, drops nulls.
• Actionable errors: a 403 suggests setting <P>_PROXY, a 401 names the auth env var, 404/429/5xx say what to do next.
• Auth: bearer, header, query, cookie or basic, read from an env var.

Comparison

Most generators mirror every endpoint as a tool and hand the model raw JSON. toolsmelt designs a few workflow-level tools with plain argument names, descriptions that say when to call what next, and compact output.

Compiled from each project's public docs in October 2026; check them before relying on a cell. Stainless, which generated MCP servers in code mode, joined Anthropic in May 2026 and closed new signups. FastMCP's own docs say curated servers clearly beat auto-converted ones.

Extending

Add an input format, a value transform or a sign-in provider as your own installed package, without forking: toolsmelt finds them through the entry-point groups toolsmelt.importers, toolsmelt.transforms and toolsmelt.signin. Plugin importers use toolsmelt.scrub, the same privacy rules as the built-ins. See docs/extending.md.

Limitations

• No browser per call: JS challenges and captchas are out of scope. Sites that need them will fail at check.
• A tool is one request, a paginated request (paginate) or a chain of requests (steps, with for_each fan-out). Logic beyond that (branching, computed values) needs hand edits to the generated server.py, and regenerating overwrites them.
• The generated project is Python only. Servers speak stdio, or Streamable HTTP with --transport http; HTTP auth is one shared bearer token (MCP_AUTH_TOKEN), not OAuth, so put a gateway in front for per-user access.
• A HAR only contains what you did while recording; unrecorded features are not in the plan.
• Endpoints read from app code are unverified until probe; a method that cannot be read stays empty.
• Request bodies are JSON, form (application/x-www-form-urlencoded) or multipart text fields (body_encoding on a tool); file uploads are not supported.
• The plan is trusted input: its text ends up in server instructions and the skill, and toolsmelt check runs the generated code. Review a plan written by an LLM from third-party content before you generate and publish it.
• Alpha.

Develop

uv sync
uv run pytest -q
uv run ruff check .

See CONTRIBUTING.md and docs/design.md.

License

MIT