Claude
Claude Desktop and Claude web can connect through the Mission MCP route. Local and hosted setup paths are available now.
Mission for LLM surfaces · Claude and ChatGPT connections
LLMs earn permission before they act.
Mission has Claude setup paths and a working hosted connector for ChatGPT. OpenAI directory approval and publication are still pending.
The same Trust Graduation boundary can sit under Claude, ChatGPT, Gemini, or another model client: prepare work freely, gate external actions by action class, evidence, approval, and receipts.
60-second proof
The ceremony is the product.
Claude asks to send an email. Mission holds it with the approval ceremony. The operator approves. The receipt is written. This demo uses the Claude route; the free OpenAI connector does not send email.
00-15sClaude prepares a specific email and asks to send it.
15-30sMission returns an approval packet: what is gated, why it is gated, and how to approve.
30-45sThe operator approves the action from Mission.
45-60sMission shows the receipt, trust delta, and boundary that prevented silent sending.
MCP RegistryListed as io.github.gomission/mcp, with public repo github.com/gomission/mcp.
Supported LLM surfaces
One Mission account. Multiple model brains.
Mission is not tied to one assistant. The model can be Claude, ChatGPT, or another LLM surface; Mission remains the account, connector, approval, and receipt layer underneath.
ChatGPT / OpenAI
The free 19-tool connector reads workspace context, captures and updates local tasks, prepares review packets and displays result cards within granted permissions. External sends, posts, Calendar changes and payments are excluded. OpenAI directory approval and publication remain pending. Review the current contract.
Gemini and other LLMs
Future model clients can use the same workspace login, connector setup link, and Trust Graduation gates as the protocol expands.
Install path
One command. Mission picks the right mode for your setup.
These install steps are for Claude and @gomission/mcp. Compatible Mission runtimes can return workspace-scoped setup links for Gmail, Calendar, Drive, X, Instagram or LinkedIn. The separate free OpenAI connector supports Mission workspace reads, task changes and preparation; it does not connect those providers or execute their actions. See its 19-tool review journey for the supported workflow.
01Run npx -y @gomission/mcp install-claude. Mission inspects your Claude Desktop config and auto-picks --wrap (gates every other MCP server automatically) if you already have Gmail/Slack/Notion/etc. installed, or --local (approval ceremony for Mission's own tools) if you don't. Override with explicit flags any time.
02Restart Claude Desktop. Ask Claude: mission status — Mission reports the mode, paste-step posture, and any unwrapped MCP servers it could be gating.
03If Mission installed in --local mode: paste the gating instruction below into Claude Desktop. --wrap mode users skip this step — gating is automatic per tool call.
04Ask Claude: set up my workspace or connect Gmail, X, and LinkedIn. Mission returns the connector setup path
Only for --local mode
Local-mode users: tell Claude to actually use Mission.
Skip this if Mission installed in --wrap mode — wrap intercepts every tool call automatically. In --local mode, installing Mission makes the gate available; Claude does not call it by default. Paste this instruction into Claude Desktop's Settings → Profile → Personal preferences, or at the top of any Mission conversation:
Mission is my permission layer. Before any external action (send email, post publicly, schedule, spend, publish, modify external records), call the request_approval tool with action_class and a one-line summary, then call mission_check_approval until decision=approved. Do not act until approved. If denied, stop and tell me. If unsure which action_class applies, call mission_classify first.
Test it: ask Claude "send an email to me at <your address> saying hello from Mission". Claude should call request_approval first; you should see the approval ceremony in the response with a receipt URL. Click the URL, approve, and Claude resumes.
If nothing happens after 3+ sessions of trying, Mission's mission_status tool will warn that the paste step is likely missing. Easiest fix: re-install with npx -y @gomission/mcp install-claude --wrap to skip the paste step entirely.
Why this is needed: Claude has no built-in mapping of "this is consequential, ask Mission first". The instruction tells Claude that mapping for the rest of the session. Anthropic's Skill directory will eventually carry this natively; until then, this is the manual wire — or use --wrap mode for structural enforcement.
What Claude can do
Preparation is fast. Consequences are gated.
Read and prepare
Find context, summarize work, rank next moves, and prepare drafts for review.
Stage external work
Draft emails, posts, packets, or submissions only after Mission exposes the approval gate.
Send, spend, promise
No external send, account change, spend, public claim, or reputation-sensitive action without explicit approval.
Trust Graduation
No LLM should need blanket permission to be useful.
Mission gives Claude, ChatGPT, Gemini, and future agent brains a visible permission layer: evidence before action, approval at the boundary, receipt after the outcome. The Trust Graduation Protocol is open. Mission is the reference implementation. Agent builders implementing their own runtimes can wrap any tool call with the @gomission/trust-graduation contract.