Olbrain Alchemist

[Olbrain]

Alchemist

OLBRAIN · AGENT BUILDER

Build, test and publish Olbrain AI agents from Claude Code, Codex, Antigravity, Cursor, GitHub Copilot or claude.ai.
You describe the agent. Claude plans it with Alchemist's guidance; Alchemist builds it, tests it, checks every step — and asks you before anything goes live, unless you let Claude decide.

[Claude Code plugin] [Codex] [Antigravity plugin] [Cursor plugin] [GitHub Copilot plugin] [claude.ai connector] [MCP: streamable HTTP] [mcptry]

Install · How a build works · Capabilities · Tools · Privacy · Docs

What it is

Alchemist is Olbrain's agent builder, available as an MCP server and a Claude Code plugin. You tell Claude what the agent should do, in plain words. Claude (or Codex, Antigravity or Cursor) plans and designs the agent from your words and your project, with Alchemist's step-by-step guidance. Alchemist builds it in your Olbrain organisation, tests it on real inputs, checks each step in order, refusing one that is skipped, and publishes it.

Every agent it builds runs on Olbrain and stays visible in Noesis, Olbrain's console, where your team can open, review and change it.

[Diagram of Olbrain's First Agent team: WhatsApp and website chat reach one front door, which asks four specialist agents: Pre-Sales, Investor Relations, Media Relations and Technical Support]
_{Olbrain's own front door, built with Alchemist: one team of agents on WhatsApp and website chat.}

Install

Claude Code

/plugin marketplace add Olbrain/alchemist-plugin/plugin install alchemist@alchemist

Then run /mcp, pick alchemist and sign in with your Olbrain account. New to Olbrain? Choose Create account on the sign-in page.

Claude Code asks before every tool call. To stop it asking each time Claude checks on a build, allow these five tools once under /permissions (they read; alchemist_state also marks the build's news as seen, and Claude shows it to you): mcp__plugin_alchemist_alchemist__whoami, mcp__plugin_alchemist_alchemist__alchemist_builds, mcp__plugin_alchemist_alchemist__alchemist_build, mcp__plugin_alchemist_alchemist__alchemist_capabilities and mcp__plugin_alchemist_alchemist__alchemist_state. Leave the rest asking, including alchemist_do and alchemist_answer: they take build steps, answer cards or change your organisation.

To see an agent drawn in a pane beside the chat, also install the canvas:

/plugin install alchemist-canvas@alchemist

It draws the agent whenever Claude shows it, and redraws it when a publish or make-live card arrives, with new, changed and removed parts marked. /agent-canvas opens the pane when the terminal is too narrow for it to open by itself. It needs a Claude Code build with plugin panes; older builds and Codex skip it.

Codex

codex mcp add alchemist --url https://alchemist.olbrain.com/mcp/codex mcp login alchemist

Optional: copy plugins/alchemist/skills/alchemist to ~/.agents/skills/alchemist so Codex follows the same workflow as Claude, including when to continue a build and when to start a new one. Without it, the server's own instructions carry the same rule. A copied skill does not update itself: copy it again after each release (0.9.0 changes how builds run).

Antigravity

agy plugin install https://github.com/Olbrain/alchemist-plugin/tree/main/plugins/alchemist

Or in the Antigravity app: Customizations → Plugins → Install from URL with the same URL. This installs the Alchemist skills and connects the MCP server. Sign in with your Olbrain account when Antigravity asks; if it doesn't, open /plugin, expand alchemist under Installed and choose Sign in. The canvas is Claude Code only.

Cursor

Install Olbrain Alchemist from the Cursor plugin directory: cursor.directory/plugins/alchemist. Sign in with your Olbrain account when Cursor asks, or from the alchemist row in Customize. A team admin can also add this repository in Dashboard → Plugins & MCPs → Import from Repo (https://github.com/Olbrain/alchemist-plugin). The canvas is Claude Code only.

GitHub Copilot

copilot plugin marketplace add Olbrain/alchemist-plugincopilot plugin install alchemist@alchemist

This installs the Alchemist skills and connects the MCP server in Copilot CLI. Ask Copilot to use Alchemist and sign in with your Olbrain account when the browser opens. The canvas is Claude Code only.

claude.ai and Claude Code on the web

Add a custom connector with the URL https://alchemist.olbrain.com/mcp/ (Customize → Connectors → Add custom connector; on Team and Enterprise plans an owner adds it under Organization settings → Connectors). Then click Connect and sign in.

How a build works

Alchemist gives one stage at a time and refuses a step taken out of order. Builds started in Noesis, and plugin builds while Olbrain moves accounts over, are still planned by Alchemist's own model through the alchemist tool; the skill switches by itself.

On a build Claude plans, each agent has a budget: six passes per fix round, a stop after three passes in a row without a gain, four failed compiles and 120 draft test runs. When a build reaches one, Claude shows the best result so far and asks whether to continue (two more passes). A stronger model usually gets a build through its tests first time.

Before Alchemist publishes an agent, makes it live, merges agents or changes how one behaves, it stops with an approval card. You decide how many of these Claude answers for you: tell it in the chat, or put it in your CLAUDE.md, AGENTS.md or GEMINI.md, for example "Approve plans and tests yourself; ask me before publishing." If you haven't said, Claude answers every card itself in auto mode (Codex: running without approvals; Antigravity: set to proceed without asking) and asks you otherwise. It still asks for what only you can do, like sharing a Drive folder or a browser step. On a build Claude plans, removing an agent, forgetting a customer, renaming or moving an agent, changing a job's email address, and publishing despite failing tests are always answered in the browser, from the link Claude gives you. If Claude Code blocks an answer, answer the card on its build page, or allow the tool with /permissions: mcp__plugin_alchemist_alchemist__alchemist_answer when installed as the plugin, mcp__claude_ai_Alchemist__alchemist_answer when added as a claude.ai connector.

Use it

• Ask in plain words: "Use Alchemist to build an agent that answers questions about our pricing and hands refunds to a person."
• /alchemist:build <what the agent should do> starts a build.
• /alchemist:status shows your account, organisation, builds and any cards waiting on you.
• "Pick up my support agent build": Claude finds your build by name and carries on. Builds are private to you; a teammate changes your published agent in a new build of their own. Anything waiting on you comes first, and Claude doesn't start a duplicate.
• "Put my support agent on my website" or "call my intake job when our signup form is sent": Claude adds the chat snippet, or a server route that calls the agent, to your project. It never puts your API key in browser code.
• "Show me my support agent": Claude draws it as a read-only diagram (example below).
• "What can Alchemist do?" or "Can it answer calls on our phone number?": Claude checks the live capability list before saying yes or no.
• "Where is my ticket TKT-…?" or "Raise a ticket: our webhook job didn't run": Alchemist handles your Olbrain support tickets in the same session.

[Diagram of the Myra chat agent: it hears from website chat, WhatsApp and Slack, checks its guardrails, consults its knowledge file and replies]
_{A chat agent's diagram: where it hears from, what it checks, what it consults and how it ends.}

Best practices

Also on docs.olbrain.com.

• Say who it serves and where it stops. Name who the agent talks to, what it does, what it must never do and when it hands over to a person. Add a few questions real customers ask: they become its tests.

  text

  Use Alchemist to build a chat agent for our website visitors. It answers pricing anddelivery questions from our FAQ, never promises discounts, and hands refunds to[email protected]. Customers often ask: "Do you deliver on Sundays?" and"Can I change my order after paying?"

• Build from your project. Start Claude Code or Codex in the folder that holds your docs, site or code. Your coding agent drafts the brief, plan and tests from it, and asks you only for what it can't find.

• Decide which cards Claude answers. Put the rule in CLAUDE.md, AGENTS.md or GEMINI.md (see How a build works). Removing an agent, forgetting a customer and publishing despite failing tests are always answered in the browser.

• Continue a build; don't start another. Say "pick up my support agent build"; a card waiting on you comes first. still_running is not an error: don't send the request again. To change a published agent, ask for the change: customers keep the live version until the new one is tested and published.

• Ask before ruling something out. "Can Alchemist …?" checks the live capability list. Charts in the chat, PowerPoint, PDF, Word and Excel files, and a job's multi-sheet Excel workbook are built in. What really can't be built yet, Alchemist reports to the Olbrain team itself.

• Give a job a real input. Share one real example of what starts it (a webhook body, an email, a file): its live test runs on it before you publish. Say whether a list is one piece of work or many: a report or workbook from a whole list is one run, each invoice in an email is a run of its own. A job that does nothing with its input fails its test.

• Let the tests drive fixes. Fix what a failing test shows, then test again. Don't start a new build to reset a budget; a stronger model usually gets through first time.

• Keep keys out of the chat. Cards collect keys, passwords and signing secrets in the browser. If one is pasted into the chat, rotate it. Keep your Olbrain Access Key on your server, never in browser code.

• Read errors by their code, not their words (Error codes). service_unavailable is an outage on Olbrain's side: the build is kept, so wait, then say "continue". insufficient_credits: top up, then continue.

• Review it once it's live. "Review how my support agent did this week": Alchemist reads real conversations, groups what went wrong, turns failures into tests and proves a fix on them before you publish it.

Capabilities

Ask Claude "What can Alchemist do?" for the full, current list: it reads alchemist_capabilities live, so it is never out of date.

You don't pick a type of agent. Describe the work; Alchemist decides.

A job starts on a schedule, a webhook (Stripe, Shopify, GitHub and other signed senders), an email, a team's front door, or a new file in a Google Drive, cloud storage bucket, Dropbox or OneDrive folder. Agents answer on website chat, WhatsApp, Shopify storefront chat, your phone number and Microsoft Teams. They use tools from Olbrain's connector catalog, any REST API or your own MCP server, and answer from their knowledge files. Every agent gets a setup review and tests before it publishes. Anything Alchemist can't build yet goes to the Olbrain team as a request, with an id you can follow up on.

Commands and tools

Inviting people and joining an organisation happen in Noesis. Every build can also be opened there.

Error codes

A failed tool call has isError set. Its structuredContent, and its text as JSON, is one envelope:

{"error": {"code": "service_unavailable", "message": "Alchemist is unavailable right now. …", "retryable": true, "reference": "ab12cd34"}}

Branch on code. Show message to the user. reference (on Olbrain failures) finds the call in our logs.

The same names are used inside results that succeed as calls:

• A build's errors[] ({code, message}): service_unavailable, insufficient_credits (with topup_url), too_many_turns (three builds already running), stream_dropped (follow the build with alchemist_build; the turn may still be running), stream_failed.
• A step's op (alchemist_do, alchemist_answer): status is done, refused or busy. A refusal's result is {ok: false, error, hint}: error is the code and hint says what to fix. Retry once after fixing it, never unchanged. provider_unavailable means the same as service_unavailable.
• Statuses, not errors: still_running and queued (follow with alchemist_build wait), busy (wait, then send the same call with the same op_id), already_started (that request_id already made the build).

Privacy and security

• Every tool call goes to https://alchemist.olbrain.com, which acts as you in Olbrain. It keeps your Olbrain sign-in encrypted so you stay signed in; remove the connection to end it.
• Keys and passwords for tools never pass through the chat. Approval cards collect them into your organisation's vault.
• Nothing goes live, merges or changes behaviour without an approval card you answered. Claude Code also asks your permission before sending each answer.
• Draft test runs, probes and compiles are counted per developer per day, to keep builds within their budget. Pricing does not change.
• Once an agent is live, its go-live steps call Olbrain's agent API at https://webhook.olbrain.com with your organisation's API key — to test it, upload a file for a job, start a job, or from the server route Claude adds to your app. You run those commands yourself, and the key stays in your own environment.

Olbrain's privacy policy and terms.

Links

olbrain.com · Noesis · Docs · Developers · Support

Licence

This plugin (its skills, commands and manifests) is MIT licensed. The Alchemist service it connects to at alchemist.olbrain.com is not part of it: using the service is covered by Olbrain's terms.