Skip to content

Integration Status

Last updated: 2026-09-20

Two different questions hide behind "do you support X?", so this page answers them separately:

  • Does the installer wire it for you? npx @misaka-net/misakanet-setup writes five agents' configs and can re-check them (--verify). Everything else is a manual recipe.
  • Can it talk to us at all? The endpoint is standard MCP over Streamable HTTP, so a client that supports MCP usually needs one config entry. That is per-vendor knowledge, and it is where the mistakes happen: the config key differs between products (mcpServers, servers, mcp, mcp_servers, mcp.servers) and so does the field that holds the URL (url, httpUrl, serverUrl).

Evidence levels used below โ€” the useful distinction is who checked it, and how:

Mark Meaning
โœ… verified A command and its output are recorded in this repository (the installer's --verify, or the field-report matrix). Re-run the command rather than trusting this page
๐ŸŸก recipe Steps are documented here, but no run is recorded โ€” treat as "we believe this works"
๐Ÿ”ต vendor-only The vendor's own docs say it works (checked 2026-09-20, links below). Nothing in this repository documents or tests it
โšช no coverage Nobody has looked

MCP Registries

Registry Status URL
Glama โœ… Listed, MCP indexed https://glama.ai/mcp/servers/Ikalus1988/MisakaNet
MCP Registry โœ… Listed https://modelcontextprotocol.io
MCP Toplist โœ… Badge live https://mcptoplist.com/server/io.github.Ikalus1988%2Fmisakanet

Agents

Agent Installer writes it Docs here MCP + remote HTTP endpoint Evidence
Claude Code โœ… ~/.claude.json, ~/.claude/CLAUDE.md, hooks + read-only tool grants in ~/.claude/settings.json claude-code.md, playbook yes, native http โœ… verified (installer + field report: 7 tools, called, lesson quoted back)
Codex โœ… ~/.codex/config.toml (mcp_servers), ~/.codex/AGENTS.md field report ยง2.4 yes, streamable-http โœ… verified (live session: config โ†’ prompt โ†’ tool call โ†’ answer)
Hermes Agent (Nous Research) โœ… ~/.hermes/config.yaml (mcp_servers), ~/.hermes/.env, ~/.hermes/SOUL.md field report only yes: url, Streamable HTTP by default โœ… verified (7 tools registered; the one no_match was a corpus bug we then fixed)
OpenClaw โœ… ~/.openclaw/openclaw.json (mcp.servers), ~/.openclaw/workspace/AGENTS.md field report only yes: transport: "streamable-http" | "sse" โœ… verified (toolSummary: {calls: 3, failures: 0})
Codewhale โœ… ~/.codewhale/mcp.json field report only yes: Streamable HTTP, legacy SSE fallback โœ… verified (codewhale mcp tools โ†’ 7; search โ†’ get_lesson)
DeepSeek Harness via the plugin channel, not the installer dsh.md yes (plugin bundle) โœ… verified (dsh --dump-config, install)
Cursor โœ… ~/.cursor/mcp.json (MCP only โ€” no rules block, no hook) cursor.md, failure-memory rule yes since 0.45: ~/.cursor/mcp.json + .cursor/mcp.json ๐ŸŸก recipe: the installer writes the entry (both installers, 2026-09-20), still nobody has run the live check โ€” that is #1940
Continue.dev โŒ continue.md per vendor docs ๐ŸŸก recipe
Gemini CLI โŒ โ€” (this page only) yes: ~/.gemini/settings.json, remote field httpUrl ๐Ÿ”ต vendor-only (mcp-server)
Windsurf โŒ โ€” yes, but read the caveat: ~/.codeium/windsurf/mcp_config.json applies to the legacy Cascade agent; the current docs moved to docs.devin.ai (MCP) ๐Ÿ”ต vendor-only, and partly stale
GitHub Copilot โœ… Copilot CLI only โ€” ~/.copilot/mcp-config.json (MCP only). VS Code and the github.com coding agent are still manual โ€” yes, in three shapes: VS Code uses servers in .vscode/mcp.json / profile mcp.json (ref); Copilot CLI uses mcpServers in ~/.copilot/mcp-config.json; the github.com coding agent is configured in repo Settings, no file ๐Ÿ”ต vendor-only
OpenCode โœ… ~/.config/opencode/opencode.json (MCP only) โ€” yes: ~/.config/opencode/opencode.json, key mcp.<name> with type: "remote" (docs) ๐Ÿ”ต vendor-only
Kiro (AWS) โœ… ~/.kiro/settings/mcp.json (MCP only) โ€” yes since launch (2025-07-14): ~/.kiro/settings/mcp.json (mcpServers), project .kiro/settings/mcp.json (docs) ๐Ÿ”ต vendor-only
Trae (ByteDance; docs now call it TraeCode) โŒ โ€” yes since v1.3.1 (2025-04-22), project scope only: .trae/mcp.json (mcpServers), gated by "Enable Project MCP". No user/global path is documented (docs) ๐Ÿ”ต vendor-only
Google Antigravity โŒ โ€” yes, one central file shared by the app, IDE and CLI: ~/.gemini/config/mcp_config.json, key mcpServers with serverUrl (codelab) ๐Ÿ”ต vendor-only, incomplete: headers, project scope, rules file and hooks are not found in reachable Google sources, not confirmed absent
Oh-my-Pi (omp) โŒ โ€” yes: project .omp/mcp.json, user ~/.omp/agent/mcp.json, type: "http" (docs) ๐Ÿ”ต vendor-only
Pi โŒ โ€” no native MCP โ€” the client is rules-only unless you add the third-party pi-mcp-adapter. Pi reads AGENTS.md/CLAUDE.md, so the failure-memory playbook is the usable path ๐Ÿ”ต vendor-only (absence of a documented MCP entry point)
Aider / Cline / Zed โŒ โ€” not checked this round; the old "Planned" rows here overstated nothing but promised nothing either โšช no coverage

The key-name trap, in one place. mcpServers (Claude Code, Cursor, Gemini CLI, Windsurf, Kiro, Trae, Copilot CLI) ยท servers (VS Code Copilot) ยท mcp (OpenCode) ยท mcp_servers (Codex TOML, Hermes YAML) ยท mcp.servers (OpenClaw). For the URL field: url is most common, Gemini CLI also accepts httpUrl, Antigravity requires serverUrl. A config entry copied from one product into another usually fails silently โ€” the server simply never appears.

Names that collide. oh-my-pi on npm is not the 32k-star agent (that is @oh-my-pi/pi-coding-agent); "Hermes Agent" is Nous Research's agent, a different product from their Hermes LLMs; codewhale is Hmbown/Codewhale (the CodeWhaleIDE repos are unrelated); Trae's IDE docs are now headed "TraeCode" while TraeWork is a separate product, and the CN build keeps rules under ~/.trae-cn/.

Codex: what "verified" means here

Codex is the one row with a full chain recorded, so it doubles as the template for the others (codex-cli 0.154.0, 2026-09-15 โ€” re-run these, don't take the table's word for it):

codex mcp list              # โ†’ misakanet | https://misakanet.org/mcp | enabled | Bearer token
codex doctor                # โ†’ config.toml parse ok ยท MCP servers 1 ยท 1 streamable_http ยท 0 disabled
codex debug prompt-input "" # โ†’ a `# AGENTS.md instructions` item carrying the rule block

The third one is the interesting one: it renders what the model will actually see, so it proves the user-level ~/.codex/AGENTS.md block reaches the prompt rather than merely existing on disk. Codex has no user-level lifecycle hook (0.154.0 hooks are admin-managed through requirements.toml), so the round-20 checkpoint reminder is rule-driven there, not hook-driven.

A live session calls the tool (2026-09-15, codex-cli 0.154.0 driven by MiniMax-M3):

MINIMAX_API_KEY=โ€ฆ codex exec --json \
  -c model_provider="minimax" \
  -c 'model_providers.minimax={name="MiniMax",base_url="https://api.minimax.chat/v1",
       env_key="MINIMAX_API_KEY",wire_api="responses"}' \
  -m MiniMax-M3 --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox \
  "pip install ไธ€็›ด timeout๏ผŒๅธฎๆˆ‘็œ‹็œ‹ๆ˜ฏไป€ไนˆๅŽŸๅ› ๏ผŸ"

โ†’ the JSONL event stream contains server=misakanet tool=misakanet_search with {"query":"pip install timeout","detail":"summary","top":5}, and the returned results carry the lesson text. The chain works end to end: config โ†’ prompt instructions โ†’ tool call โ†’ answer. Two notes for anyone re-running it: Codex 0.154.0 rejects wire_api = "chat" (Responses API only โ€” MiniMax serves /v1/responses, so "responses" is the working value), and --json is what makes the tool call visible (plain exec output does not name it).

Other

Integration Status Notes
Smithery โธ Paused Pending upstream changes
Any other MCP client โœ… unverified but standard https://misakanet.org/mcp speaks Streamable HTTP and answers tools/list with 7 tools; reads are anonymous and unmetered. If your client needs a bridge (mcp-remote, supergateway), say so in #1550 and we will document it

Field reports

  • Agent integration matrix โ€” 2026-09-15/16 โ€” one row per agent (Claude Code, hermes, openclaw, codex, codewhale, DSH), the command that produced the evidence, and what each agent actually did with MisakaNet. Re-run the commands in it rather than trusting this table.