Cursor, OpenCode and other MCP clients

Cursor

kontext connect cursor --write     # writes .cursor/mcp.json

{ "mcpServers": { "kontext": { "command": "kontext", "args": ["mcp"] } } }

Cursor also reads .cursorrules; kontext's brief lists it under Rules for agents. kontext connect agents-md --write adds the working rules to AGENTS.md.

Hooks. kontext connect cursor --hooks --write writes .cursor/hooks.json (version: 1): sessionStart and postToolUse (after a file edit), both running kontext hook --host cursor. Any --write also adds beforeShellExecution, which denies kontext trust run by the agent. Cursor also reads Claude Code hooks from .claude/settings.json; install one set, not both. Opt-in until the A/B report; see Hooks.

Skills. Cursor reads SKILL.md files from .claude/skills/, where kontext connect skills --write puts them.

OpenCode

kontext connect opencode --write   # writes opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": { "kontext": { "type": "local", "command": ["kontext", "mcp"], "enabled": true } }
}

OpenCode reads AGENTS.md and the skills in .claude/skills/. It has no hook protocol kontext writes to; the agent gets the brief by calling ctx_brief, as the AGENTS.md block tells it.

Superset and other orchestrators

Orchestrators that run many agents in parallel worktrees need nothing special: each agent's harness starts its own kontext mcp in its worktree. All worktrees of a clone share the inbox, the outbox and bootstrap progress (they live in the git common directory); each worktree keeps its own search index. State of worktrees that no longer exist is pruned.

Two things to know with orchestrated agents:

• Fresh clones have no hooks. Run kontext hooks install once in a project the orchestrator cloned (worktrees share the hooks). Until then the brief warns, and ctx_prepare_commit says the trailers have to be added by hand.
• Agents leave the session's directory for other worktrees, submodules and sibling projects. See Several repositories.

A useful division of labour:

• one agent runs /mcp__kontext__kontext-init to deepen the bootstrap while others work,
• every agent calls ctx_brief with its focus paths,
• captures from all agents meet in the shared inbox, and each agent promotes the ones belonging to its own change with ctx_prepare_commit.

Any other MCP client

Configure a stdio server with the command kontext and the argument mcp, started in the repository directory. If the client starts servers elsewhere, pass the repository explicitly:

{ "command": "kontext", "args": ["-C", "/path/to/repo", "mcp"] }

or set KONTEXT_DIR=/path/to/repo in the server's environment.

Clients without MCP

Use the CLI from any harness that can run shell commands:

kontext brief --focus src/api
kontext search "retry policy" --json
kontext call ctx_prepare_commit '{}'