Skip to content
ollama-intern-mcp

ollama_chat

{/* PER-TOOL REFERENCE TEMPLATE Each tool page in this directory follows the same shape: 1. One-line job description (name what intern-job this names) 2. Tier + when-to-use 3. Schema — derived from src/tools/.ts (zod) 4. Example call + envelope 5. Common pitfalls 6. Related tools Keep it under ~200 lines. The full tier-grouped overview lives at ../. */}

ollama_chat is the catch-all for ad-hoc model interaction. It is visibly last-resort by design. If you find yourself reaching for chat more than once a session, a specialty tool is missing — file a feature request rather than rely on chat as your normal entrypoint.

Tier — Workhorse

Section titled “Tier — Workhorse”

hermes3:8b on the default dev-rtx5080 profile. The Workhorse tier shape budget is ~4–8k tokens per turn (per TierConfig.num_ctx for your profile).

When to use it

Section titled “When to use it”
  • One-off prose generation that doesn’t fit any specialty tool
  • Light back-and-forth where structured output would be overkill
  • Bridge calls where a future feature will replace this with a job-shaped tool

When NOT to use it

Section titled “When NOT to use it”
  • Classification → ollama_classify (gives you confidence + threshold + frame)
  • Extraction → ollama_extract (schema-constrained JSON, no parse errors)
  • Reading a corpus → ollama_corpus_answer (chunk-grounded citations)
  • Reading specific files → ollama_research (path-grounded citations)
  • Producing an artifact → packs (incident_pack / repo_pack / change_pack)

Schema

Section titled “Schema”
{
  messages: Array<{
    role: "system" | "user" | "assistant";
    content: string;       // min 1 char
  }>;                      // min 1 message
  system?: string;         // optional preface, merged with any system messages
  model?: string;          // per-call model override (advanced — use sparingly)
}

Full source: src/tools/chat.ts.

Example call

Section titled “Example call”
{
  "tool": "ollama_chat",
  "arguments": {
    "messages": [
      { "role": "user", "content": "Summarize this commit message in one sentence: 'feat(corpus): surface :latest drift on refresh'" }
    ]
  }
}

Returns:

{
  "result": {
    "reply": "Adds drift detection to corpus refresh when the source path uses ':latest' tag resolution."
  },
  "tier_used": "workhorse",
  "model": "hermes3:8b",
  "hardware_profile": "dev-rtx5080",
  "tokens_in": 38,
  "tokens_out": 24,
  "elapsed_ms": 920,
  "residency": { "in_vram": true, "evicted": false }
}

Common pitfalls

Section titled “Common pitfalls”

You’re using chat to extract structured data. Switch to ollama_extract with a schema. chat reply is a free-form string — you have to parse it yourself, and the parse will fail half the time.

You’re using chat to classify. Switch to ollama_classify. It gives you a confidence score, an allow_none escape, and a threshold floor. chat does not.

You’re chaining 5+ chat turns. That’s a sign the job is bigger than chat should handle — consider splitting into a corpus_index + corpus_answer flow, or a repo_brief if you’re trying to characterize something.

You’re hitting context limits. chat runs at the Workhorse tier’s num_ctx. For longer-form prose, reach for ollama_summarize_deep (Deep tier, larger window).

Section titled “Related tools”