MCP & skills
NomiFun has two easily-confused extension paths. This page helps you tell them apart and set them up:
- An MCP server is an external tool server that exposes callable tools over stdio / HTTP / SSE, giving the agent a batch of new “hands-on” abilities.
- A skill is a markdown / folder knowledge pack that tells the agent how to complete a workflow — it is knowledge, not a long-running tool server.
In the other direction, NomiFun’s own 151 capabilities are exposed through three public fronts — /mcp (full Remote), /mcp-agent (a curated do-work subset), and /v1 (REST + OpenAPI 3.1 + SSE streaming) — so Claude / Codex / any agent can drive NomiFun directly via MCP/Skill.
Entry points: the sidebar MCP page (
/mcp) configures external servers, the Skills page (/assistants?tab=skills) manages general skill packs, and outbound exposure lives under/open-capabilities. Old Settings URLs redirect to these pages.
Connect an external MCP server
-
Add a server. Open
/mcp, click add, name it, and pick a transport — stdio takes command / args / env, HTTP / SSE takes a URL. Importing from another agent’s config preserves its raw JSON too.
-
Test the connection. Before saving, run the connection test: the backend starts a temporary MCP client, completes the handshake, lists tools, and persists the result. Failure codes distinguish missing command, permission, timeout, HTTP, RPC, and protocol errors, which makes debugging easier.
-
Handle OAuth. HTTP / SSE servers that need OAuth go through the built-in OAuth flow; once authorized they connect.
-
Import / sync agent configs. Detect MCP configs from supported local agent CLIs and import detected servers into NomiFun in one click; when the adapter supports writing, you can also sync NomiFun’s MCP list back into a selected agent’s config. This step is configuration management only.
-
Select per session. Enabling globally only makes a server available — it does not auto-inject into every session. Which servers a given session actually sees is the merge of three sets: globally enabled servers + servers selected for that session + the builtin bridge servers the current capability set requires.
Manage skills (general knowledge packs)
Open /assistants?tab=skills to manage skill packs shared across companions. A skill can be a single markdown file or a directory containing a SKILL.md, and is sourced one of three ways:
| Source | Meaning |
|---|---|
| Builtin | Ships with the app; some are auto-injected into sessions |
| Custom | Skills you import or drop into the config directory |
| Extension | Provided by an installed extension |
Skills can be tagged, imported, exported / symlinked, scanned from external directories, and materialized for a specific agent backend.
Two kinds of “skill” — don’t mix them up
What you manage at /assistants?tab=skills are cross-companion, general skill knowledge packs. That is a different thing from a “companion’s own skill library”:
- General skills (this page) — knowledge packs you curate by hand and every companion can share, sourced as Builtin / Custom / Extension.
- Per-companion skill library (on the companion page) — each companion grows its own, per-companion-isolated set of skills. The evolution engine mines reusable skills from that companion’s tool-call sequences (based on the sequence itself only, with no parameter values), produces a draft, and you review the resulting
SKILL.mdbefore enabling it; grown skills can also be gifted to other companions. See Companions.
Keep the memory-vs-skill boundary straight: memory is a shared memory hub, shared family-wide across one capture / learning pipeline; skill libraries are isolated per companion. Per-companion private memory is coming soon.
Exposing NomiFun to external agents
Manage outbound fronts under /open-capabilities. Authentication is by companion-token (Bearer, SHA-256 stored, constant-time compared); the caller operates “as that companion” on the Remote surface. The headline capability is nomi_agent_run (streaming) + nomi_agent_result (polling) delegation. Review each front before enabling it.
Notes & boundaries
- Enabled globally ≠ injected into a session. Enabling an MCP server globally only makes it “available”; whether it enters a given session is decided by that session’s selection.
- The connection test is a temporary client. The test only handshakes, lists tools, and stores the result; it does not affect later real connections in sessions.
- The evolution engine sees sequences, not parameters. Per-companion skill mining is based on tool-call sequences only and never records parameter values, so the distilled skills are safe to review and gift.
- Current version is 0.1.0 (pre-1.0). Review some fronts (such as the open-capabilities panel) before turning them on.
Related
- Companions — per-companion skill libraries, evolution mining, and skill gifting.
- Knowledge base — attach a knowledge base to a session / terminal / companion so the agent can “look it up first”.
- Terminal — where the agent actually runs commands.
Full docs → GitHub