Skip to content

armgit / Ix

Public Imported from GitHub
0

Understand any codebase instantly. System intelligence for codebases, built for humans and AI.

5.6 KB
Raw Blame

Consolidating the per-host plugins onto ix mcp

Ix ships six per-host plugin repos, each with its own tool implementation: ix-claude-plugin, ix-codex-plugin, ix-cursor-plugin, ix-gemini-plugin, ix-openclaw-plugin, ix-opencode-plugin. ix mcp serves the same graph tools from the CLI, so those implementations can be retired in favour of one server.

This is what that costs and in what order it can happen.

What MCP can and cannot absorb

Measured across the six repos (source only, excluding node_modules and dist):

Layer Lines Absorbed by ix mcp?
Tool implementations (mcp/, tools/, runtime/) ~15,300 Yes
Agents, hooks, skills, commands, rules ~10,600 No

MCP standardizes tools, resources and prompts. It has no concept of a hook, a skill, a slash command or a host instruction file, so roughly 40% of what the plugins do has no MCP equivalent and stays host-specific whatever else happens. Consolidation retires the tool layer; it does not retire the plugins.

Tool-surface parity

The plugins never shared one tool surface. Measured against ix mcp:

Plugin Tools Tools ix mcp lacked
ix-codex-plugin 23 —
ix-cursor-plugin 23 —
ix-gemini-plugin 23 ix_decide, ix_ingest, ix_query, ix_status
ix-openclaw-plugin 17 ix_decide, ix_docs_tool, ix_ingest, ix_neighbors, ix_query
ix-opencode-plugin 17 ix_decide, ix_docs_tool, ix_ingest, ix_neighbors, ix_query
ix-claude-plugin 0 (skills only) —

ix mcp now closes that gap, with three deliberate exclusions:

  • ix_query — the deprecated command the CLI docs tell agents not to use ("produces oversized low-signal responses"). Dropped on purpose.
  • ix_neighbors — a composite of ix_callers + ix_callees + ix_depends, all served individually. Callers lose one round-trip, not a capability.
  • ix_docs_tool — a composite over ix_overview, likewise served.

And ix_status is Gemini's name for what ix mcp calls ix_health; both run ix status.

ix_decide and ix_ingest were real gaps and are now served. ix_decide is a write and needs Pro, so it is advertised only when @ix/pro resolves — as are ix_briefing and ix_decisions, which this server previously offered to every install even though every call answered "requires Ix Pro".

Net: codex and cursor can delegate with no change in what an agent can do. gemini, openclaw and opencode lose only ix_query and the two composites.

Structured results and tool annotations

ix mcp exposes two protocol conveniences on top of the tool surface:

  • Tool annotations. Every advertised tool carries a title, and every tool implemented in this repository also carries readOnlyHint, destructiveHint, idempotentHint and openWorldHint so a client can present and route it correctly. Graph reads are read-only and idempotent; ix_map, ix_ingest, and ix_smells in its default mode, which stores versioned claims, are destructive because they mutate backend state; only ix_ingest is open-world, because its GitHub form reaches an external API.

    The Pro tools — ix_briefing, ix_decisions, ix_decide — carry a title and no behavioral hint, for the same reason they carry no outputSchema: they are implemented in the separate @ix/pro package, so nothing here can check a claim about them. That asymmetry is deliberate rather than an oversight. Clients use readOnlyHint to decide whether a call needs the user's approval, so a wrong one does not merely mislabel a tool — it removes the prompt in front of it. An absent hint costs a confirmation; a wrong one costs the confirmation itself.

    Hints describe intent and are not themselves a boundary: the enforcement that matters (argv construction, timeouts, output bounds, process cleanup) lives in the runner, and what a tool can reach is bounded by the command it runs, not by what it is advertised as.

  • Structured content. ix_map, ix_ingest and ix_smells return their parsed JSON object as structuredContent (with an outputSchema) so agents receive typed values instead of a string to re-parse. The schema asserts only "an object": the real field shape is backend-defined and versioned, and promising specific fields would break tool calls the moment the backend renames one. The human-readable text result is preserved for clients that prefer it. Most remaining tools return their llm-format text, which has no stable object shape worth promising.

    The Pro tools are the exception worth naming: they run in JSON too, but get neither an outputSchema nor structuredContent, because their shape cannot be checked from this package. Those two go together on purpose — attaching the parsed object without the schema would hand clients the unverified value anyway, minus the contract that would let them validate it. Their answer is still returned, as text.

Order of work

  1. Ship ix mcp in a released @ix/cli. Nothing downstream can depend on it until then. The plugins already gate features on CLI version floors, so this becomes a new floor.
  2. Point each plugin's registration at ix mcp instead of its own server, and delete that server. Start with codex and cursor — those two are lossless.
  3. Leave agents, hooks, skills and commands where they are. They are the plugins' remaining reason to exist.

Step 2 is one PR per repo and each is independently revertible; a plugin that has not migrated keeps working, because ix mcp install refuses to take a name that another server already holds.