Set up your agent

Connect your own agent to your existing project. Kitecraft supplies the plan and researched writing packets; your agent writes in your tools. For a supported hosted chat with ChatGPT or Claude and no local setup, see Connect with hosted chat.

Connect a local agent in three steps

The steps below are for local coding agents with a Bearer developer key. They do not apply to hosted chat OAuth connections.

Open Connect your agent and create a read-only developer key. The key is shown only once. Use a write key later if you want the agent to edit your plan or request research.
Supply the key privately as KITECRAFT_API_KEY in your agent's environment or secret store. Never paste it into a chat or commit it to a repository. The agent process must inherit that environment.
Copy the handoff below into your agent. It configures a remote MCP connection, verifies your account and asks which project to use when you have several.

Give this to your agent

Provide your key privately as KITECRAFT_API_KEY, then paste this handoff into your agent. It connects and reads your existing project before proposing work.

MCP agent handoff

Connect my agent to Kitecraft and resume my existing project.

Read https://kitecraft.ai/docs/mcp for connection details and tool schemas.

1. Add a remote Streamable HTTP MCP server named kitecraft at https://kitecraft.ai/mcp. Use the Bearer token from KITECRAFT_API_KEY in my environment or secret store. Never print the key or put its literal value in prompts, source files, URLs or committed configuration. Preserve my other MCP servers. Use this client's supported secret substitution. Kitecraft developer keys do not use OAuth. If a key or client configuration is missing, tell me exactly what I must supply. If this client cannot use a remote server with a Bearer header, explain that limitation and offer the documented CLI connection instead. For an explicitly supplied preview/local origin, use that origin with /mcp appended.

2. Initialize the connection and discover its tools. Call project_context and verify the returned account, MCP server contract version and projects. Check that its advertised calendar_handoff, calendar_setup and artifact_save/list/read tools exist in this host, and brief_request declares recoverOperationId. If a tool or argument is missing, refresh the supported host connection and repeat discovery. A local agent can use the documented authenticated CLI/API with its existing private key; a hosted chat must refresh its connection. Report the precise mismatch instead of claiming the new workflow passed. Never ask me to paste credentials into chat. Verify the returned account and projects. Use my existing project. If several projects exist, ask which one I mean. Retain its project ID for every project-scoped call. Do not recreate my account, restart my trial or overwrite saved setup. If no project exists, report the next setup action from onboarding_status before changing anything.

3. Read calendar_handoff for a compact queue, then calendar_handoff with one returned itemId. Inspect the current status and next action. Preserve existing reviewed work and retrieve its exact saved article/review with artifact_list and artifact_read. Full profile_read, content_list and plan_read archives are explicit supporting reads for a particular missing fact, not default startup context. When I ask you to run one saved article workflow, use calendar_run and follow its bounded instructions, stopping for my review. For recurring setup use calendar_setup after verifying a single bounded run; creating or reading setup does not enable a schedule.

4. For an eligible writing item, read the brief_read overview, then its writer input and exact source/material reads. Follow nextOffset until null for required text, pin identities and verify the complete content hash. A summary is not sufficient writing evidence. Follow the exact free recovery call returned by brief_status; do not replace it with a fresh purchase. Ask before missing/stale research or hosted evaluation. Retain the operation ID and poll with a deadline; queued work is unfinished. Save completed article/review bytes with artifact_save, read back their hashes and report current work using returned revisions. A local path alone cannot be read by a cloud agent.

5. Return the verified project, connection result, next article, packet readiness and one clear next action. When I ask you to write, follow the packet's draft and edit instructions in my own tools. Flag unsupported claims. Do not publish, start research, enable schedules, change billing or delete data without my instruction.

Prove the connection with an actual project_context response. A saved MCP configuration alone is not a successful connection.

Your API key is never included.

What success looks like

Your agent reports an actual project_context result, the intended project, its next article and writing-packet readiness. A configuration file or a “connected” label by itself does not prove the agent can read your project. Existing setup is reused, and reading the saved packet does not start new research.

If a brief is missing or stale, the agent explains that before requesting research. Research uses your allowance and needs your instruction. Checkout and provider consent stay in your browser.

Client connection details

The production endpoint is https://kitecraft.ai/mcp, using Streamable HTTP with Authorization: Bearer authentication. This is a hosted server; you do not need to install or run a Kitecraft MCP server locally. Developer keys do not use OAuth.

Each agent client has its own configuration format and secret handling. Your agent should follow its client's documented remote-MCP configuration and preserve your other servers. Clients that cannot send a Bearer header can use the CLI instead. Use the complete Codex desktop or CLI setup or OpenCode setup. Codex desktop must actually inherit the secret environment variable or use its private header helper; a terminal export alone does not prove GUI access.

When the packet is ready, ask your agent to draft from it, follow its editing instructions and flag unsupported claims before you review or publish. See Writing briefs for the complete packet workflow.

Optional destination setup

For WordPress.com, select the credential-free agent connection. Connect your agent to the official WordPress.com MCP, confirm the intended site and enable only the required draft/media tools. Use Kitecraft preparation and reviewed artifacts, then WordPress tools to upload, save a private draft and read it back. Kitecraft stores references, not image files or WordPress.com tokens. See WordPress.com setup for Codex, Claude and ChatGPT instructions.

Open Article destination for the intended project and read destination_status. For Custom API or self-hosted WordPress, your agent inspects the supported API and prepares a reusable non-secret mapping; you enter the credential only in the masked owner form. Save the selected destination, verify its read access, then use destination_prepare for a no-send package. Retain the article and package, obtain independent review and approve its exact bytes before an authorized send. Article destinations gives the complete human and agent setup, downloadable examples and recovery steps.

For Squarespace, save the site without credentials and skip native API verification and reads. Use destination_prepare to create the editor handoff, retain it, complete independent review and obtain owner approval before editing. Save and reopen the private draft in the separately authorized Squarespace editor; no Commerce API key or native send/readback is involved.

Month planning and calendar execution

Start with project_context, then calendar_handoff for a compact queue and one selected item. Preserve and retrieve existing draft/review bytes before creating replacements. Use calendar_setup for maintained recurring instructions, after verifying one bounded run. Use calendar_month and calendar_month_resume only for an explicitly requested new horizon. calendar_read, calendar_claim, calendar_commission and calendar_report support both bulk preparation and customer-scheduled wakeups. Read the complete workflow for exact transitions, owner approval, CMS readiness and recovery. Saving preferences enables no recurrence. Download the current writing workflow skill; previous frozen skill versions remain available.

Verify the installed contract

Compare the MCP server version and tools from fresh discovery with the workflow advertised by project_context. It includes calendar_handoff, calendar_run, calendar_setup, artifact_save, artifact_list, artifact_read and destination operations; brief_request includes recoverOperationId. Reusable mappings also require destination_prepare. Missing names or constraints can mean an outdated host connection or a feature not yet deployed there. Follow Refresh a stale connection, rediscover the actual tools and schemas, then verify again. A local agent may use the authenticated CLI/API with its existing private key. Do not request keys in chat, create another key just to diagnose discovery, or call a backend route proof of native host discovery.