[INTEROP] Export Registry agents to MCP and major code runners

[EmersonBraun]

Goal

Make one source-owned AgentsKit Registry agent installable and runnable across major coding-agent hosts without maintaining one implementation per host.

Parent interoperability charter: AgentsKit-io/agentskit#1477

Current baseline

• npx agentskit add <id> copies source into the consumer project.
• Each Registry entry already has agent.ts, meta.json, tests, README, and evaluation evidence.
• The first universal projection should be MCP; native host projections remain thin generated artifacts.

Phased delivery

1. Extend the existing metadata contract only where needed for projections: input/output schemas, capabilities, secrets, network, side effects, approval, target support, provenance, and evaluation evidence.
2. Export one read-mostly agent (recommended: coding-code-qa or another verified code-review agent) as a namespaced MCP tool.
3. Pin exact versions and integrity hashes; add clean-environment discovery/invocation/error/timeout/approval tests.
4. Verify the MCP export with DeepSeek Harness and a second MCP host.
5. Generate Codex, Claude, and Cursor skill/rules projections with fail-closed host detection.
6. Verify Gemini, Kimi, Grok, Hermes, and other hosts only through documented public MCP/skill/plugin/CLI contracts.
7. Add “Add to…” snippets, compatibility state, security evidence, and cross-links to every Registry agent page.

Safety constraints

• Never silently install or execute latest.
• Declare secrets, network, tools, side effects, and human approval.
• Default to read-only; require explicit approval for destructive operations.
• Do not add host-specific dependencies to @agentskit/core.
• Do not mark a host supported without an executable smoke test.

Acceptance criteria

• One agent remains source-owned after install.
• The same release exports to MCP reproducibly.
• MCP discovery and invocation pass in two clean hosts.
• Codex, Claude, Cursor, DeepSeek, Gemini, Kimi, Grok, and Hermes are individually marked verified, partial, planned, or unsupported.
• Registry pages expose canonical install/export snippets and evidence links.

Broader backlog (provider swaps, memory/RAG, Doc Bridge, production integrations, Vercel AI/LangChain/CopilotKit coexistence, badges/SEO, and demand-driven native plugins) is intentionally tracked as later phases rather than mixed into the first vertical slice.

[EmersonBraun]

Pilot correction: start with the validated research agent for the generic MCP vertical slice. It has one skill plus default fetch/search tools and no caller-injected command runner, so the existing @agentskit/mcp --agents wrapper can represent it honestly. Multi-stage agents such as code-review should follow only after a dedicated projection contract for their source-specific inputs is defined.

[EmersonBraun]

Phase 0/first vertical slice is implemented locally: runner projection statuses are now validated and emitted for the research pilot, the generated bundle carries the pinned source-install command and MCP metadata, and the README includes a pinned @agentskit/mcp@0.3.9 recipe. The MCP state remains partial until a clean-host execution smoke test is recorded; Codex/Claude/Cursor/DeepSeek/Gemini/Kimi/Grok/Hermes remain planned rather than being overstated. Focused projection/validation tests, registry validation, lint, build, ecosystem eval replay, and Doc Bridge gate pass. The full test suite still has the pre-existing registry-root: sourceHash is stale freshness failure in readme-standard.

[EmersonBraun]

Implementation update: the first published MCP smoke test passed with @agentskit/mcp@0.3.8 and the Registry research agent. The stdio server initialized and exposed fetch_url, research, and web_search; filesystem and shell were not enabled. The public recipe is now pinned to 0.3.8 because 0.3.9 is not published on npm (an attempted 0.3.9 install correctly failed with ETARGET). The projection remains partial until host-specific approval UX is verified.

[EmersonBraun]

Reproducibility tightened: the Registry pilot now records the exact npm tarball integrity for @agentskit/mcp@0.3.8 (sha512-qY3c128bLrvdsqPOeTDYg/Ian4yhK3HR1eOdPb/pqIrGlHDckPa04p9JivzPOx2yYYc1/sfrur+gh3Ftl6gSKg==) alongside the version. Generated public/r/research.json, index.json, and catalog.json were rebuilt from the canonical metadata.

[EmersonBraun]

Next slice added to the canonical research README: copyable MCP previews for Codex, Claude Code, and Cursor, all pinned to @agentskit/mcp@0.3.8 with the Registry agent selected and shell/filesystem disabled. They are explicitly labeled planned until each host's approval and invocation path is smoke-tested.

[EmersonBraun]

Host evidence round: Codex accepted the pinned command through its read-only config override (codex mcp get), and Claude Code wrote/validated the expected project .mcp.json in an isolated temporary project. Cursor's CLI accepted the MCP payload but its installed build did not expose the documented workspace flag, so no Cursor support claim was upgraded. All three remain planned until a real host tool invocation is captured; the published MCP process itself remains verified at the protocol/tool-discovery level.

[EmersonBraun]

The host ledger now records the evidence without overstating support: Codex configuration parse accepted; Claude Code project config accepted but claude mcp get/list correctly reports the server as pending approval; Cursor remains pending because its installed CLI lacks the workspace flag. No host was promoted to verified or partial from configuration-only evidence.

[EmersonBraun]

DeepSeek Harness projection added to the canonical research README. It uses the official @deepseek-ai/dsh-mcp-client stdio contract with serverName: agentskit, pins @agentskit/mcp@0.3.8, and documents the expected namespaced tools (mcp__agentskit__fetch_url, mcp__agentskit__research, mcp__agentskit__web_search). It remains planned until a clean Harness invocation is run.

[EmersonBraun]

Second-wave discovery smoke: the published bridge can initialize and list tools for research-academic-synthesizer, research-industry-benchmark, and content-fact-checker. We are not promoting them yet: these agents rely on invokeStructured, Zod schemas, and agent-specific safety/output logic that the generic Registry-to-MCP wrapper intentionally does not execute. The correct next slice is a dedicated typed-agent projection contract, not a misleading generic task -> text export.

[EmersonBraun]

Phase 1 follow-up: the projection contract now fails closed on agent semantics.

• projections.<host>.mode accepts skill or typed.
• typed projections require an object outputSchema; optional inputSchema, when present, must also be an object.
• The published research projection is explicitly mode: skill because the generic MCP wrapper preserves its skill prompt and tool surface.
• Agents that rely on invokeStructured, Zod output schemas, or agent-specific safety nets remain unprojected until a source-backed typed executor preserves those semantics; discovery alone is not enough.

Registry validation/build and focused projection tests pass. Next implementation slice: choose one typed agent, publish its source schema/executor contract, then run discovery + invocation/error/timeout smoke in MCP and a second host before upgrading its status.

[EmersonBraun]

Next slice started with research-academic-synthesizer as the first typed candidate. Its canonical manifest now publishes an explicit object input schema and the invokeStructured output schema (including requiresReview: true), while the MCP projection remains planned/typed.

The source executor test passes, but no generic MCP export was promoted: the typed adapter must preserve structured output and human-review semantics before the status can change. Registry build, validation, focused source/projection tests (11), and docs bridge gate pass.

[EmersonBraun]

Interop guard added in @agentskit/mcp: hosted and raw Registry manifests with projections.mcp.mode: typed are rejected by the generic skill fetcher, so structured agents cannot be accidentally exposed as task -> text tools. The MCP package suite passes (48 tests) and TypeScript lint is green.

[EmersonBraun]

The AgentsKit MCP bridge now has a typed projection adapter (createTypedAgentTool) and a separate fetchAgent path. It validates the declared input/output schemas, preserves structured results, and keeps fetchAgentSkill fail-closed for typed manifests. The Registry research-academic-synthesizer candidate remains planned until the MCP package release and host-level smoke evidence are available.

[EmersonBraun]

The typed MCP projection is now consumable by the published adapter: @agentskit/mcp@0.4.2.

• Integrity: sha512-PzCxQwAGOySk9IDgFaoosbqETkswVAEyUuE2TdVVg9E4zRyUwq7tGRTTuzHSSg8zBmvShPpBOQNZn1brnx3FiA==
• npx -y @agentskit/mcp@0.4.2 --help passed from outside the monorepo.
• research-academic-synthesizer is marked mcp: partial; host-specific approval and invocation smoke tests remain before verified.

Registry artifacts were rebuilt and validation/docs bridge gates pass.