| name | use-gotcontext |
| description | Route large context through gotcontext.ai MCP compression before sending to a model — pre-flight, ingest, skeleton reads, PR diffs, remote docs. Use when about to paste a large file, verbose CLI output, git diff, or remote URL into context. Pair with tg for local code reads. Triggers include "compress this", "shrink context", "gc_pre_flight", "ingest_context", "gotcontext MCP", "token savings". |
use-gotcontext
gotcontext-saddle is the behavior harness; gotcontext.ai is the compression
engine. This skill routes you to the right layer.
Install (once per machine)
- API key — sign in at gotcontext.ai/dashboard/settings,
create a
gc_-prefixed key.
- Export before every session (MCP reads the shell, not
.env.local):
export GOTCONTEXT_API_KEY="gc_..."
- Wire MCP without replacing existing config — pick one:
- Saddle installer (additive merge):
./install.sh claude (or codex, cursor, agy)
- Official CLI:
npx --yes gotcontext@1.0.1 wrap claude (or codex / gemini)
Do not copy a template over an existing user configuration.
- Restart and verify:
- Claude / Codex / Gemini CLI:
npx --yes gotcontext@1.0.1 doctor
- Cursor: open Settings → MCP and confirm
gotcontext is enabled
- agy: run
/mcp and confirm gotcontext is connected
- Optional plugin (outcome-shaped compression skills):
/plugin marketplace add oimiragieo/gotcontext-sdk
/plugin install gotcontext
Plugin skills: shrink-for-claude, pre-flight, review-pr-diff,
compress-mcp-manifest, ingest-docs, batch-compress, extract-api-surface,
session-summary. Read SETUP.md in the plugin on first run.
Routing table
| Task | Tool | Notes |
|---|
| Large pasted doc / handoff | ingest_context → read_skeleton | profile=core; give a unique file_id |
| Remote URL / docs page | ingest_context(file_url=…) → read_skeleton | profile=core; avoid raw full-page fetches |
| Find a section in ingested doc | search_semantic → modulate_region | profile=core; evidence-aware reads |
git diff, pytest -v, CI logs | filter_cli_output | profile=full; use when output is genuinely verbose |
| Route a large payload | gc_pre_flight or plugin pre-flight | profile=full, Pro+ |
| PR review / changed symbol context | gc_blast_radius or plugin review-pr-diff | profile=full, Pro+; submit snippets |
| MCP tools/list bloat | gc_compress_manifest or plugin compress-mcp-manifest | profile=full; shrink tool descriptions |
| Local repo search | tg search PATTERN PATH | Scoped; broad scans refused |
| Local symbol before edit | tg callers PATH SYM / tg blast-radius | Disk primitive |
| Server-side code navigation | gc_callers, gc_blast_radius, … | profile=full, Pro+; submit snippets |
PreToolUse hooks enforce this (installed with ./install.sh claude):
Read ≥20 KB full-file → blocked → use MCP, scoped Read, or tgWebFetch → nudge ingest+skeletonGrep / shell grep/rg → nudge tg searchPreCompact → core compression path; gc_pre_flight only when profile=full is enabled
Emergency: export SADDLE_GOTCONTEXT_BYPASS=1
MCP profiles
?profile=core (installed default) — compact core compression profile with a
small tool-list footprint. It does not expose gc_pre_flight or the
server-side code-navigation suite.?profile=full — complete catalog, including Pro+ routing, code navigation,
security, KB, and batch tools. Your account plan still controls access.
Change the url in your MCP config or pass the profile query param.
When NOT to use
- Secrets, tokens, credentials, raw auth output
- Prompts under ~200 tokens (send them as-is; compression overhead is not worthwhile)
- Already-compact green test output
- Sources you must reproduce character-for-character
Full discipline write-up
core/prompts/principles/08-gotcontext-dogfood.md
Docs
