gopher: fetch, cache and digest context for Claude

Three things, one install

fetch

A whole repository as markdown: directory tree plus the files that matter. Binaries, lock files and node_modules filtered out, the rest ranked.

cache

Memory that survives the conversation, stored as a structured context.json and an append-only diary.md. Both readable without this tool.

digest

Reads a transcript, pulls the facts out with a local Ollama model, and merges them into the cache. Nothing leaves the machine.

[ THE DIGEST ]

Most of a repository is not worth sending

Early versions returned a fixed ten files, which is not a size limit at all: ten files is anywhere from four thousand characters to two hundred thousand. Anything past the client's ceiling was rejected whole, so the caller got nothing rather than something trimmed.

Before

27,194  test/ecosystem/airflow/pyproject.toml
32,048  test/ecosystem/home-assistant/...
27,517  test/ecosystem/pandas/pyproject.toml
 9,156  test/ecosystem/jupyterlab/...
 8,757  test/ecosystem/black/pyproject.toml

228,310characters, mostly the wrong files

After

Cargo.toml
README.md
pyproject.toml
Dockerfile
crates/uv-build/src/main.rs

39,168characters, within budget

There is a total character budget now, spent on the tree first and then files in priority order until it runs out. Ranking accounts for where a file sits, not just what it is called, so a root manifest beats a vendored copy of the same filename buried in a test fixture.

Repository	Before	After
astral-sh/uv	207,867	39,321
modelcontextprotocol/python-sdk	217,488	39,224
punkpeye/awesome-mcp-servers	228,310	39,168
sktime/sktime-mcp	103,786	39,231

[ MEMORY ]

Memory you can only read in full is a file

Facts accumulate across conversations. Reading them back used to mean returning the entire store every time, which works until it does not. Search over key paths and values, or read one section by its dot path.

Reading the whole store

read_context()

13,484characters returned

Finding one fact

search_context("memory")

202characters returned

{
  "query": "memory",
  "matches": [
    {
      "key": "projects.precedent.note",
      "value": "agentic memory hackathon, submitted",
      "matched": "value"
    }
  ],
  "shown": 1,
  "total": 1
}

Substring matching over keys and values, no embeddings and no index. On a store of a few hundred facts that is enough, and something cleverer can wait for evidence that it is not.

[ QUICK START ]

Running it

Clone and install

git clone https://github.com/pyarchana/gopher.git
cd gopher
uv venv
uv pip install -e ".[dev]"

Optional: a GitHub token, for a higher rate limit

cp .env.example .env
# add GITHUB_TOKEN=... to that file

claude mcp add gopher --scope user -- \
  uv run --no-sync --project /absolute/path/to/gopher gopher

Or with Claude Desktop, in claude_desktop_config.json

{
  "mcpServers": {
    "gopher": {
      "command": "uv",
      "args": ["run", "--no-sync", "--project",
               "/absolute/path/to/gopher", "gopher"]
    }
  }
}

The digest tools additionally want Ollama running locally with llama3.2 pulled. Fetch and cache work without it.

[ TOOLS ]

Ten tools

Tool	What it does
`fetch_github_repo(repo_url)`	Markdown digest of a public repo, tree plus top files
`search_context(query, limit)`	Find facts whose key or value contains the query
`read_context(prefix)`	The whole store, or just one section
`update_context(key, value)`	Set a value at a dot path
`delete_context_key(key)`	Delete a key by dot path
`log_diary(entry, tag)`	Append a timestamped markdown entry
`read_diary(last_n)`	Read back the last N entries
`digest_transcript(transcript_path)`	Extract facts from a transcript, merge into the store
`summarize_only(transcript_path)`	Same extraction, returns JSON without writing
`read_digest_log()`	Full contents of the digest log

[ CONFIGURATION ]

Four environment variables

Read from a .env file in the project root, which the server loads on startup.

Variable	Default	Purpose
`GITHUB_TOKEN`	none	Raises the GitHub rate limit from 60 to 5,000 an hour
`GOPHER_DATA_DIR`	./data	Where the store, diary and digest log live
`GOPHER_DIGEST_BUDGET`	40000	Maximum characters one repo digest may return
`GOPHER_OLLAMA_MODEL`	llama3.2	Model used for fact extraction