Changelog

A tool call now waits at most 30 seconds for your API, then answers with an error your client can retry, instead of hanging for as long as the client would wait. Servers get this once they're redeployed.

When a GraphQL API rejects a key inside a normal 200 response, as many do, a server that uses each caller's own key now asks for a new one, as it does for a 401, and errors no longer say the call "failed with status 200". Errors beside data that still came back are passed to the agent instead of dropped. Connection fields such as GitHub's are asked for their first 10 items, since those APIs refuse them without a page size. Servers get this once they're redeployed.

sokie init with a GraphQL introspection file says GraphQL is made in the app, rather than that it isn't valid OpenAPI. sokie deploy says when another deploy of the server is still running. sokie dev says to set SOKIE_CALLER_KEY when a tool needs your key. For a GraphQL server, sokie check and sokie tools --unexposed no longer suggest x-mcp: false, which a schema can't say, and a warning about a type that leads back to itself calls it that, not the depth limit.

A server's name can now be up to 64 characters in the app, as in sokie.json, so a server made on the New server page keeps its whole name in its files. The New server page also says .gql files can be dropped, as they always could.

Settings › Workspace › Webhooks sends a signed POST to your own systems when a deploy goes live or fails, a server takes its first call, usage passes your alert or calls stop, or a server's sign-in to its API breaks. Each webhook gets a signing secret, shown once, and failed deliveries are tried again for half an hour. Send a test from Settings, and see the last week's deliveries under each webhook. The docs have a Webhooks page with how to check the signature.

Settings has a Notifications tab with a switch for each email about your servers: a first call, a failed deploy from GitHub, a broken sign-in to your API, and usage passing your alert. Each of those emails also has a link to turn it off in one click, no sign-in needed. Sign-in links, billing and plan emails, and tool calls stopping always come.

The docs have a search box over the page list: type a few words and pick a result to land on the section that answers. New pages cover GraphQL APIs (where the schema comes from, what each tool asks for, and what it can't do yet), Troubleshooting (what a refused call, a stopped deploy or a stuck sign-in means, in the words Sokie shows) and Limits and plans (every limit you can reach, and what raises it).

The New server page and the sample API button said they couldn't reach Sokie when your plan's servers were all in use. They now say that, and how to make room: delete a server on its Settings tab, or upgrade on Billing.

Every dialog that asks you to confirm shows its keys: Esc on the button that backs out, ⌘↵ (Ctrl+Enter on Windows and Linux) on the one that goes ahead, and both work from anywhere in the dialog. Where a dialog asks you to type a server's name, the name stands apart and copies with a click.

Access, cases, deploy keys and deleting a server move from the bottom of the Overview to the server's new Settings tab (press 7), each card with its own mark. The Overview is now just how the server is doing.

The Deploys tab has a Redeploy button beside production. It builds the live deploy's code again on the newest runtime, and production keeps answering until the new deploy is live. When a newer runtime is out than the one your server runs, the tab says so, so runtime fixes like the Logs tab's call names reach it without the CLI.

A client connecting to your server makes a few calls before its first tool call: initialize, notifications/initialized and tools/list. The Logs tab, Copy logs and sokie logs now show those by name where they used to show no tool, and a refused call with no name says it came before the client signed in. Servers say this once they're redeployed.

After choosing a plan in Stripe, Billing says it's waiting for Stripe to confirm your move to Pro or pay as you go, then that you're on it the moment Stripe does, usually within seconds. If Stripe takes longer than it should, it says who to write to.

The Connect panel now adds your server to Claude Code, Cursor and VS Code under its name, like weather-open-meteo, rather than its random address, so it reads well in the client's server list, in Logs, and in the first-call email.

When Claude, Claude Code, ChatGPT, Cursor or GitHub Copilot connects to a server, and on the front page, the badge on Sokie is now that app's own mark instead of one Sokie drew.

No GitHub or Google account you want to use? Type your address on the sign-in page and Sokie emails you a link that signs you in, and makes your account if you're new. The link works once, for 10 minutes, and opens a Sokie page where you press Sign in, so a mail scanner that opens links can't use it up. An address gets at most five links an hour, and Sokie still keeps no passwords.

Before your first server, the Dashboard asks for your spec's address and opens the New server page already reading it. No spec to hand? Try it with a sample API makes a weather server on Open-Meteo's free API, which needs no key, so you can call it straight away. The CLI's three commands are still there, folded under the card. The welcome email points to the same place.

Once a server is live, its Overview shows how to add it to claude.ai, ChatGPT, Claude Code, Cursor and VS Code, with its address filled in: the steps, the command to copy, or a one-click install for Cursor and VS Code.

The first time an agent calls one of your server's tools from its own client (not the Test tab or sokie call), Sokie emails you once to say which client called which tool, and the server's Overview shows it for a week with what's worth doing next: reading the call in Logs and checking the tools agents see.

When your API takes a key from each caller, sokie init now asks for yours (hidden as you type) and saves it in .sokie/dev.env, which git ignores, so sokie dev can call your API straight away. Press Enter to skip; run from a script, it says how to add the key later instead.

A server made on the New server page now says its deploys were made in Sokie from a spec, not uploaded from the CLI, on its Overview, its Deploys tab and its GitHub card. Servers made before today keep the old words.

A tool calling an API that sends its answers deflate-compressed, as Open-Meteo does, got unreadable bytes instead of JSON; it now gets the JSON. And a spec that names its server on a path or an operation rather than at the top, as Open-Meteo's does, now finds its address in sokie init, sokie check and the New server page.

A new band on the front page shows the New server page's three steps, from a spec's address to a live server, and what it took for Swagger's Petstore: under a minute.

Mail apps that darken an email in dark mode leave its images alone, so Sokie's black tile disappeared into the dark and only its face showed. Sokie in emails now has a thin light edge that keeps the tile visible, and is invisible on light mail. The waitlist email's Sokie winks now, as the welcome email's does.

While Sokie lets people in a few at a time, signing in with an account that isn't let in yet puts your verified address on a waitlist instead of turning you away: Sokie makes no account, sends one email saying you're on the list, and lets you in the same way once there's room. Every sign-in is now checked by Cloudflare Turnstile, and sign-ins from one address are limited each minute. An account whose email isn't verified with GitHub or Google is told to verify it first.

sokie.sh/security says in plain words how Sokie protects your API's credentials, your callers and your data: OAuth for every client, encrypted secrets and caller keys, logs that never keep what a call sent, the audit trail, and how to report a problem. The front page gains a band about it, and says that a server past 40 operations offers agents three tools (search, read and call) so a big API fits in an agent's context, and that each tool's risk tells clients when to ask you first. The footer links to Security and to an address to write to, and sokie.sh/llms.txt gives agents a map of the docs.

npm i -g @abyssinia-labs/sokie installs the sokie command (and sok) on Node 22 or later, under the Apache 2.0 license. Its npm page says how to go from a spec to a live server, and lists every command. A new version reaches npm only once sokie.sh runs the matching update, so the CLI never asks for something the site can't do yet.

The front page now says up front that Sokie turns an OpenAPI, Swagger or GraphQL spec into a live MCP server, and links to sokie.sh unfurl with a picture and that line instead of a bare title. On a phone, a menu reaches How it works, Your agent, Pricing and Docs. Each plan has its own button (Start free, Start Pro, Talk to us), and the three-command demo ends with a way to deploy your own spec.

sokie --version (or sok -v) prints which version of the CLI you have, so you can tell whether an update reached you. With --json it answers { "ok": true, "version": … }.

Installing the Sokie App on GitHub, changing which repositories it sees, or asking an organization's owners to approve it now brings you back to Settings › GitHub, which says what to do next. Before, GitHub left you on a page telling you to go back to your terminal. Add an account now opens GitHub in the same tab, so the trip is there and back. An account shows in the list once a server links one of its repositories.

Server logs now keep calls for as long as your plan says: 7 days on Hobby and pay as you go, 30 on Pro and a year on Scale. Before, every plan kept 7. Moving to a plan that keeps fewer days removes older calls within the hour; moving up keeps the calls still there for longer, but can't bring back ones already removed. The Logs tab and sokie logs say how far back yours go.

After moving to a plan with fewer servers, the newest servers past its limit keep serving what is live, but deploys from the CLI, the app and GitHub, and resuming a paused one, are refused until you delete servers or upgrade. The Deploys tab and sokie deploy say why, and a refused push fails its Sokie check on GitHub.

The usage warning now heads every page, a server's own page included, so a stopped server says why its calls fail. Billing's meters measure what the warning does and go red only when calls stop: pay as you go and Pro no longer look stopped past their included calls, and a spending cap has a meter of its own. Pro with a cap warns again near the cap. Moving your alert or cap and crossing it again sends one more email. The CLI shows a stop even on a day it already warned, and with --json writes the warning as a JSON line on stderr. A stop or resume that didn't reach your servers at first is put right within minutes.

Cancelling a plan in Stripe's billing portal now shows "Ends on …" on the Billing page and sends the email saying when it ends. Before, Sokie missed the way Stripe's portal schedules a cancellation.

sokie.sh now has a Terms of service and a Privacy policy, linked from the site's footer. They say in plain words what you agree to, what Sokie keeps about you and your servers, who helps run it, and how long anything stays. Both are drafts until a lawyer has reviewed them, and each page says so.

Turning on pay as you go, moving to Pro or back to Hobby sends the workspace's owner an email with the new plan and how its limits compare with the old one. Cancelling Pro sends one saying when it ends, with a link to keep it. Receipts and invoices still come from Stripe.

Billing shows your plan and this month's tool calls and servers against it. Hobby stays free with no card: 50k tool calls a month, 3 servers and 7-day logs. Turn on pay as you go to keep calling past the free calls at $2 per 100k, up to a monthly cap you set, or move to Pro at $19 a seat a month or $190 a year for 1M calls, 20 servers and 30-day logs. Stripe handles cards and invoices, and AI is never on the bill.

At 80% of the month's free tool calls, a spending cap or Pro's included calls (or a share you choose in Billing), the Dashboard, Billing and the CLI say so. A Hobby workspace without pay as you go, or one at its cap, has tool calls answered with a message saying why until the month resets; connecting still works, and turning on pay as you go or raising the cap brings calls back at once. Calls past what a plan includes are billed as they happen.

The warning at your alert threshold and the stop at a limit now reach the workspace's owner by email too, in the same words as the Dashboard, with a link to Billing. Each comes at most once a month, and nothing is sent when calls resume.

Sokie emails the workspace's owner when a production deploy from GitHub fails or times out, and when a server's shared sign-in to its API stops working, with a link to the page that fixes it. A run of failures sends one email, not many. New accounts get a short welcome with the three steps to a first server.

Sokie's emails now open with a coloured band that says what kind of news it is, red when something is broken, amber for a warning, green when things are good, with Sokie in a mood to match. The facts follow in a short table: which deploy failed and why, what production still serves, or when a shared sign-in broke. Each has one clear button, and the plain-text version says the same.

On a server's Deploys tab, Link a repository walks you through it: connect GitHub, pick an account and repository, then the folder with its sokie.json, compared with what the server runs now. The tab shows the link's health and what each push and pull request did, and Settings › GitHub lists where the Sokie App is installed. Change folder reopens the dialog on the folder step, and unlinking takes the open previews down.

New server takes a repository, a pasted spec or a URL, shows the tools it found before anything is made, lets you leave some out or fix their descriptions and effects, then deploys it. Download its files to push them yourself when you're ready.

Pick a deploy on the Deploys tab to see what agents saw differently: tools added or removed, descriptions, effects, inputs, response shape and runtime. sokie deploys diff shows the same in a terminal.

Every deploy has a number, and shows the commit and branch, pull request or CLI upload it ran. A server reports which deploy is answering on /health, sokie deploy prints the number, and sokie call and sokie logs name it.

sokie edit clones the server's repository (or finds your clone), opens it in VS Code, Cursor, Zed or your $EDITOR, and starts sokie dev. Edit locally on the server's header hands you the command.

The Overview is now a bento: whether the server is answering, calls per hour today, the most-called tools, recent deploys and previews, how agents see its tools, latency, errors, connected clients and the mix of what its tools can do.

A Dashboard for the workspace's servers and today's calls, and a Usage page for the month, by server and by tool. Search moved to the top bar.

Keep cases as tests/<name>.json in the repository. sokie test runs them, they run after every deploy, and the Test tab's Cases view shows how each did.

⌘K or Ctrl+K opens a palette for servers, pages and actions, and runs sokie commands the app can do. Shortcuts show the way your system writes them; press ? for the list.

The Audit page records who changed what: access, secrets, links, deploy keys and more, kept for a year.

Building, checking and serving a server the size of GitHub's or Stripe's API takes a fraction of the time it did, and calls no longer slow down as the tool list grows.

A JSON spec up to 14 MiB deploys, and a linked repository fetches only the files that changed. A YAML spec deploys up to 2 MiB; larger ones are asked for as JSON.

The Test tab copies what the agent got or the whole MCP exchange, and sokie logs pipes cleanly into other tools.

Sokie is the o in the wordmark, and wears its moods there.

The site's bar hides as you scroll down, the changelog is public, and the front page leads with a spec becoming a server.

Upstream calls now send a User-Agent, which GitHub requires, unless a server sets its own.

On large servers, searching a tool's exact name or title puts it first.

sokie call <tool> reaches tools on a server that lists them by search.

After a redeploy, sokie deploy says the previous deploy may answer for up to a minute while the address switches over.

See every tool as agents see it, with its effect, inputs and the operations no tool exposes. sokie tools lists them and sokie call calls one, on sokie dev or a deploy.

Call a tool on production or a preview, as you, and see what the agent gets.

On a linked server, change a tool's description, title, tags, effect or response shape, expose an operation, or fix a finding. Each save opens a pull request.

The Logs tab and sokie logs show a server's calls from the last seven days, live.

Every call is counted, and limited per caller, at the edge.

Make a deploy key on a server and your workflow can deploy it with the sokie GitHub Action, without the Sokie App.

Make a server public or private, pause and resume it, or delete it, from its page.

A quickstart, how-to guides and concepts at /docs, and a CLI reference generated from the CLI itself.

The sidebar shows your workspace, Settings is split into You and Workspace, and Billing has a page.

Sokie's site: a spec becoming a live server, the three commands to try, previews, Wake and pricing.

Servers, deploys, tools and logs load in pages, so large workspaces stay quick.

The sign-in's ground moves like water under your pointer.

When GitHub sent you straight back, sokie link reported a failure while Sokie was still checking.

Your name and picture refresh from GitHub or Google each time you sign in.

A server deployed before tools were kept rebuilds its tool list when you open the tab.

Install the Sokie GitHub App on a repository and every merge deploys the server. Each pull request gets its own preview URL, and sokie check runs as a GitHub Action.

A server whose API uses OAuth sends each caller through the provider's sign-in, so tools act as that person, not as you.

Servers belong to a workspace instead of a single person, ready for teammates.

Each deploy lists notes on tool names, descriptions and risky operations, the same ones sokie check prints. No AI involved.

sokie init, check, fmt, dev, login, deploy, secrets and config, each with --json. sokie login signs in through your browser.

Owner secrets are sealed and bound into every deploy. A caller can add their own API key on the connect page instead.

See every server, its status and endpoint, and a live history of its deploys with what failed and why.

Sokie's moods play twice and then hold still, so pages stay calm. Hover to play one again.

MCP clients sign in to your servers through Sokie's OAuth server, with your consent, and their tokens rotate.

Sokie animates through its moods across the app. Turn the animation off in Settings › Appearance.

Your name, email and picture come from the account you sign in with.