new Meet overapi, the fourth app. A powerful API workbench — and every API you bring in can become tools your agents can safely use. download ↓
codelions v0.1.0 · betaGitHubDownload

[ overapi ] · a powerful API client, built for agents

A powerful API workbench.
Ready for agents.

Bring in any API — a server on your Mac, a spec, SOAP, or a web app you record — and work with it like the best desktop clients. Then one reviewed pass makes it agent-ready, and it’s an MCP server in Claude Code or your team’s overgo. No account, no cloud; secrets stay in the keychain.

a server on this Mac · OpenAPI · Swagger · GraphQL · WSDL · HAR · Postman · cURL · docs · source code · web apps

[ ways in ] · bring it in however you can

Almost no API comes with a clean spec. Start from whatever it has.

Swagger behind a login, SOAP, a wiki page, a Postman export — or nothing but the calls a web app makes. Every way in ends as requests you can send, and any token it finds goes to the keychain, never into the project.

:4010A server on this Mac

New API lists what’s listening locally and imports a served spec in one click. Spec behind a login? Paste a token and it fetches it anyway.

:4010 Acme Orders node ✓ spec · 7
✦Read its code

No spec at all? Your own claude reads the service’s routes, request and response classes and sign-in — read-only — and drafts every endpoint.

✦ Read its code
{ }Import

OpenAPI 3, Swagger 2, GraphQL, a HAR capture, a Postman v2 collection, or a WSDL. GraphQL queries become reads and mutations writes; SOAP operations get typed arguments.

acme-orders.openapi.yaml
$_Paste a cURL

“Copy as cURL” from the network tab. Ids in the path become arguments, the values you sent become the example.

curl '…/orders/1002' -H 'Authorization: Bearer …'
From documentation

PDFs, Markdown, a docs URL or pasted wiki text. claude finds the real spec files if the site links to them, and drafts one if it doesn’t.

Draft the API
●Record a web app

Use the app in overapi’s browser — or your own Chrome, password manager and all — and its API calls become requests.

chrome · app.acme.example

tip: copy a cURL, a spec URL or a docs page and press ⌘V anywhere in the app · ⌘K jumps to any request

[ the workbench ] · use it like any API client

Fast enough to live in. Smart about ids.

What’s being sent sits on top, one line each; everything else the request takes waits below as chips. An id another request gives back is picked by name from what it answered — or taken from it fresh each time — and path values show by name right in the URL bar.

params as a query

  • A control per kind of value: any · true · false, chips for lists, a stepper for numbers
  • The address bar spells out the query as you go; paste a URL to fill the params
  • Load a past call that worked
  • Responses as a table, a chart, or the full JSON; variables suggested as you type

ids and chains

  • Ids by name — “Wile E. Coyote”, not 1 — sent as picked, or taken from the list each run
  • Chains: which request gives the ids another needs, kept for the next call
  • A login request runs first when the API needs one
  • Copy as cURL, fetch, Python requests or HTTPie — secrets as •••

environments and sign-in

  • Local, staging and production, colour-coded; new ones start read-only
  • OAuth 2 (presets for Google, Microsoft, Okta and Auth0), a token from a command like gcloud or az, API keys, a login request, a browser session — or another API’s
  • When a sign-in runs out, Sign in again redoes it the right way and resends
  • Client certificates per environment; multipart uploads, GraphQL and SOAP bodies

[ record ] · no spec? use the app

Build the API from what actually happened.

Sign in and use a web app in overapi’s browser — or your own Chrome, on overapi’s own profile so your password manager stays — and every call it makes to its backend is kept. overapi spots the sign-in, follows ids from one response into the next, and turns routes into requests with real examples.

Or press Explore for me and it clicks around on its own, reading only: it never submits a form or presses anything like buy, delete, send or sign out, and any request that isn’t a read is stopped before it leaves the browser.

Found 14 requests · 5 chains · 3 new in the last minute✦ Explore for me Exploring page 12 of 40 · 3 new · 2 stopped (not reads) https://app.acme.example/orders/1003/shipments orders list ✓ one ✓customers list ✓ one ✓shipments list ✓ one – Same importer as a HAR capture: ids become path params, values you sent become arguments. Secrets go to the keychain. Everything starts read-only.

[ agent-ready ] · one reviewed pass

A hundred cryptic tools become the ones an agent can choose by.

Point an agent at an API raw and it gets ids it can’t find and a DELETE it shouldn’t see. Make it agent-ready tidies what a recording picked up, fixes values that never change and connects the chains — then your own claude or codex writes descriptions, argument help, clearer names, what to hide, a brief for the agent and example questions.

You see every change, before → after, and the readiness and security scores move as you pick.

READY FOR AGENTS58 → 94SECURITY74 → 91Apply 11 changes
TIDY UP · 2POST /analytics/eventRemoved CHAINS · 3get_order gets orderIdlist_orders → [].id → get_order.orderId DESCRIPTIONS · 4 ✦get_orderOne order with its line itemsOne order — its line items, status and total. Find the id with list_orders first. HIDE FROM AGENTS · 1 ✦delete_orderDeletes for good — nobody should ask an agent for this. AGENT BRIEF ✦Order ids are numbers — find them with list_orders or search_orders before get_order. Totals are in US dollars.
rules first

Tidying, fixed values and chains are worked out by rules, so they’re the same every time. AI only writes what needs words.

a tool budget

Past about 40 tools, agents pick worse and pay for the list every turn. The Tools tab shows the budget and suggests what to turn off, and why.

security scan

Leaked secrets, what agents can reach, prompt injection in descriptions, personal data in examples. Critical and high findings block team publishing until fixed or accepted with a reason.

try as an agent

Ask a question in plain words. Your claude answers through this API’s own MCP server, read-only, and you watch every call — then save the run as a test.

[ tests ] · proof, not a promise

Prove it works — and keep proving it.

Every Try as an agent run can be saved as a test: the calls the agent made, in order. So can a recording of you using the app. overapi passes ids from one response into the next and checks every answer’s status and shape — so when the API changes underneath your agents, a test fails loudly before an agent fails quietly.

Run them in the app before you publish, or in CI with overapi test --junit. Writes are opt-in per step and never run on production. Load tests of reads are there too, behind a setting that’s off by default.

$ overapi test acme-orders/overapi.json --env staging --junit results.xml ✓ GET list_orders 200 88 ms ✓ GET get_order 200 39 ms ✓ GET get_customer 200 35 ms ✗ GET list_shipments 200 47 ms $.items was an array, now an object · POST cancel_order skipped — changes data 3 passed, 1 failed, 1 skipped in 0.4 s$ overapi scan acme-orders/overapi.json --fail-on high

[ safety ]

Read-only until you mean it.

Every tool is read, write or destructive — POST …/search counts as a read; DELETE and POST …/delete are destructive. Write tools stay hidden until an environment allows writes, and destructive ones also need confirm: true and carry destructiveHint, so clients ask you first.

Secrets live in the keychain and are scrubbed from everything shown or returned. A project someone else shared — a clone, a pull — can’t use your secrets until you’ve reviewed where it sends requests and how it signs in, and allowed it. Your browser sign-ins never publish.

SECURITY SCORE74C1 high2 medium HIGHLEAKED SECRETS−15get_customer's saved values holds an API tokenFix: keep it in the keychain and refer to it by name.🔒 Move to the keychainAccept… MEDIUMPERSONAL DATA−6get_customer returns personal data Found by rules, so the same API always scores the same. Run it in CI with overapi scan.
▤Acme Shop3 APIs · signed in by Acme Auth WHERE EACH ONE RUNS ● local● staging● production Acme Auth:3000auth.stg…auth.acme… Acme Orders:4010orders.stg…orders.acme… Acme Inventory:8089inv.stg…add production… ONE MCP SERVER · IDS ACROSS SERVICES list_orders → items[].sku → get_stock.skuAcme Orders → Acme Inventory

[ workspaces ] · the APIs of one product

One sign-in. One map of where they run. One MCP server.

Group the services that belong together. One of them signs the others in, one grid sets where each runs in every environment, and they’re served together as a single MCP server — with ids flowing from one service to the next.

Each stays its own API; the workspace only adds what they share.

[ publish ] · one project, every agent

One file. Claude Code, Codex, overcli — or your whole org.

Each install is its own MCP server for one environment: the same tools, that environment’s base URL, sign-in and writes setting. A running server notices edits within seconds and tells the client its tools changed — no restart. And the Publish sheet shows whether its tests pass before anything goes out.

this MacOne click into Claude Code

Adds the environment to Claude Code — and so to overcli, which uses the same list. Codex and any MCP client get a config to paste. Nothing is uploaded.

claude mcp add acme-orders-staging --scope user \
  --env API_TOKEN=<API_TOKEN> -- \
  overapi serve overapi.json --env staging
your orgHand it to overgo
  1. 1. You publish. The project goes to overcli, then to overgo — with no secret in it.
  2. 2. An admin finishes it. In overgo’s console (Setup → APIs): read-only or writes, and the secrets typed there.
  3. 3. Workers use it. overgo runs it as a tool workers can be given, like any other.
what overgo is →
a fileShare it through git

A project is a folder — overapi.json and a .gitignore that keeps .env out. That’s the shared collection: no account, no cloud workspace. Teammates open it in overapi or serve it from the command line.

overapi check overapi.json
overapi call  overapi.json list_orders \
  --env local --args '{"status":"open"}'
overapi serve overapi.json --env local

[ download ] · grab a build

Signed, notarized, and on your machine.

overapi 0.1.0 is pre-1.0, and it updates itself — on the stable channel, or beta if you’d rather. The macOS builds are Developer-ID signed and notarized by Apple, with the same Developer ID as the rest of the family, so they open on a double-click. It comes with a fictional REST API, a SOAP service and a web app to try it on.

macOS only for now. The AI steps — reading code or docs, Make it agent-ready, Try as an agent — use your own claude or codex CLI, signed in. Every build and its checksums are on the releases page.

[ faq ]

Questions people ask.

Isn’t this just an OpenAPI-to-MCP generator?

Generators are the easy part, and they need a spec. overapi starts from whatever you have — a server on your Mac, its source code, a WSDL, the docs, or a recorded web app — and spends its effort on what generators skip: one reviewed pass that names, connects and documents the tools, a budget so there aren’t too many, a security scan, writes kept from agents until you allow them, and tests from real traffic.

Is there an account or a cloud?

No. overapi is an API client that runs on your machine. Projects are JSON files on disk, secrets are in the keychain, and the MCP server it makes runs locally when Claude Code or Codex starts a session. The only copy that ever leaves is one you publish to your own org’s overgo — without secret values.

Can an agent change data by accident?

Environments start read-only, and tools that change data aren’t even listed until an environment allows writes. Allowing writes on production means typing its name. Destructive tools also need confirm: true on every call and carry the MCP destructiveHint, so clients that ask first will ask. A tool that isn’t on doesn’t exist, as far as the agent can tell.

Does the AI see my API’s data?

Only where you use it. Discovery from documentation sends your docs to the claude or codex CLI already on your machine, which reads them with file-reading tools only — no shell, no web, no MCP. It writes an OpenAPI draft that goes through the normal importer, and drafted requests are marked until you’ve sent them once. The agents you publish to see what the tools return, capped by each tool’s response limit.

What does it cost?

Nothing. Like the rest of the family it’s Apache 2.0, with no account and no paid tier. The AI features use a CLI you already have and are already signed in to.

How does it fit with overcli and overgo?

Anything you add to Claude Code from overapi shows up in overcli too, since it uses the same list. Publishing to your org goes through overcli, which signs in to your overgo; an admin approves it there and workers can be given it as a tool — the same way overdb hands over a database connection.

[ family ] · over your tools, not instead of them

codelions →