Ollama Intern MCP — Handbook
Ollama Intern MCP gives Claude Code a local intern with rules, tiers, a desk, and a filing cabinet. Claude picks the tool; the tool picks the tier (Instant / Workhorse / Deep / Embed); the tier writes a file you can open next week.
No cloud. No telemetry. No “autonomous” anything. Every call shows its work.
The shape
Section titled “The shape”Four tiers, 28 tools total.
| Tier | Count | Purpose |
|---|---|---|
| Atoms | 15 | Job-shaped primitives (classify, extract, triage_logs, summarize_*, draft, research, corpus_*, embed*, chat). Batch-capable atoms accept items: [{id, text}]. |
| Briefs | 3 | Evidence-backed structured operator briefs — incident_brief, repo_brief, change_brief. |
| Packs | 3 | Fixed-pipeline compound jobs that write durable markdown + JSON. incident_pack, repo_pack, change_pack. |
| Artifacts | 7 | Continuity surface — list, read, diff, export_to_path, plus three deterministic snippet helpers. |
Freeze lines: atoms+briefs at 18, packs at 3, artifact tier at 7. The surface does not grow.
Why this project exists
Section titled “Why this project exists”Every local-LLM MCP server leads with token-savings. Ours leads with what the intern produces:
- a durable markdown file you can open tomorrow
- an evidence block where every cited id was verified server-side
- a
weak: trueflag when the evidence doesn’t support the claim — never a smoothed narrative - investigative
next_checks, never “apply this fix”
Where to go next
Section titled “Where to go next”- Getting started — install, Claude Code config, model pulls
- Tool reference — every tool grouped by tier
- Envelope & tiers — uniform envelope, hardware profiles, residency
- Artifacts & continuity — how packs write to disk and how to use what they wrote
- Laws & guardrails — evidence-first, no remediation drift, deterministic renderers
- Security & threat model — what’s touched, what’s not, what’s in the log
- Corpora — build, refresh, search, answer over a living corpus; manifest v2 +
:latestdrift - Error codes — every structured error code, when you’ll see it, what to do
- Use with Hermes — drive this MCP from Nous Research’s Hermes Agent on hermes3:8b (validated 2026-04-19)
- Troubleshooting — Ollama not running, model pull failures, hardware insufficient, MCP server not appearing in Claude Code
- Observability — read the NDJSON log, field semantics, jq recipes, degradation signatures,
ollama_log_tail - Comparison — honest matrix vs other local-LLM MCPs, raw Ollama, and Claude-direct