Connect Tileward in about a minute.
Tileward Context keeps a long conversation small; Tileward Governance refuses an off-policy prompt before the model sees it. Both run against the same API key. Pick your client and copy four lines.
Paste it into your agent. It fetches tileward.com/agent-setup/prompt.md,
configures Tileward Context for whichever client you are in, and stops to ask you for an API key. It does not
install Governance.
Fetch and execute the appropriate instructions to set me up for Tileward from https://tileward.com/agent-setup/prompt.md
Three clients, two capabilities, one API key.
The same key drives both capabilities everywhere they are available. What differs is what each client lets a third party do, and that is a real difference rather than a roadmap gap.
| Client | Tileward Context | Tileward Governance |
|---|---|---|
| Claude Code | Yes — MCP server | Yes — prompt hook |
| OpenAI Codex | Yes — MCP server | Yes — prompt hook |
| ChatGPT | Yes — connector | No, and not planned |
Governance needs somewhere to stand in front of the model. Claude Code and Codex both run a hook before a prompt reaches the model and honour its refusal; ChatGPT has no such seam, and no third party can add one. Anyone telling you otherwise is describing a system prompt, which is a request rather than a gate.
Which address, and what it wants.
Two ways to authenticate. Pick by what your client can do, not by preference — a tool that cannot run an OAuth flow needs the key endpoint, and a tool that will not hold a static secret needs the OAuth one.
| URL | Authentication | Use it from |
|---|---|---|
https://context.tileward.com | API key, Authorization: Bearer | Claude Code, Codex, anything you configure by file |
https://context.tileward.com/mcp/connect | OAuth sign-in | ChatGPT connectors, Claude connectors |
https://api.tileward.com/v1/guard/hook | API key | Governance, when you want no local script |
https://api.tileward.com/mcp | API key | Legacy. Still served, so existing configs keep working |
The Context host takes one string with no path to append. If a client asks you for a "server URL", that whole string is the answer. Get an API key from the console under Settings → API keys; Governance also needs a policy bound to that key, because a key with no policy refuses nothing.
The shortest thing that works.
Context needs no plugin, no marketplace and no local server. It is a URL and a key.
claude mcp add --transport http tileward-context https://context.tileward.com \
--header "Authorization: Bearer $TILEWARD_API_KEY"
# OpenAI Codex — ~/.codex/config.toml
[mcp_servers.tileward_context]
url = "https://context.tileward.com"
bearer_token_env_var = "TILEWARD_API_KEY"
# ChatGPT — developer mode, add a custom connector pointed at
https://context.tileward.com/mcp/connect
Then ask the model to remember something and recall it in a later turn. /mcp in Claude Code
or Codex lists the tools once the server connects.
What is not available, and why.
Tool-call governance
Governance gates the prompt. Gating each tool call the agent then makes — refusing rm -rf /
while allowing ls — is built on the client side and not on ours: the endpoint does not exist yet, and a
key has no execution policy to bind. The client refuses to pretend otherwise, so it is shipped unregistered rather
than failing open and looking installed.
Conversation scoping
Context isolates on the conversation argument that each tool call carries, and in an MCP client the
model passes it. It is instructed to and normally does, but nothing enforces it, so a turn that omits it
lands in a shared per-key store where a later recall can surface another thread's material. Verified against
production on 2026-08-26: with the argument, a second conversation could not see the first's material; without it,
it could. Use one key per project where that matters.
Governance in chat clients
ChatGPT, Claude Desktop and every other chat surface have no pre-prompt hook. Governance is available where a client will run a gate before the model sees the text, which today means coding agents.
What leaves your machine
Context receives the turns you ask it to remember and your recall queries — it cannot return a slice of something it was never given. Governance receives the prompt text being checked. Neither is used for training.