Managing MCP across agents: compare your options by the numbers
August 25, 2026 · View on GitHub
Four ways to give multiple AI agents the same MCP tools — what each one actually costs you in setup, tokens, and safety. Numbers first, opinions second. Companion pages: workflow before/after · benchmark methodology
The four approaches
| Manual per-agent configs | Config-sync CLIs | GUI managers | mcptoon | |
|---|---|---|---|---|
| Add one server, N agents | edit N files × N formats | 1 command, syncs | 1 form, syncs | 1 command (mcptoon add + auto-detect) |
| Agent-side setup required | yes (per agent) | yes (per agent) | yes (per agent) | no — any shell-capable agent works day one |
| Platforms | — | usually macOS/Linux-first | often macOS-only (e.g. commercial managers) | Windows / macOS / Linux equal (pure Python stdlib) |
| Dependencies to install | n/a | Node/npm trees common | bundled runtimes | zero (~250KB) |
| Token cost of tool discovery | full JSON schemas | full JSON schemas | full JSON schemas | −88.5% (slim) or −99.8% (name-only manifest) |
| Result payload size | raw JSON | raw JSON | raw JSON | −34% (TOON encoding) |
| Security inspection of results | none | none | none | injection / credential-leak / destructive-op guards on every call |
| Health checks across agents | manual | some | some | built-in (mcptoon health, CI exit codes) |
The capability table is deliberately category-level: it compares approaches, not individual products. Products change; trade-offs don't.
What discovery actually costs (measured)
Every turn, an agent that supports MCP loads tool definitions into context.
Same 255 tools, four encodings, token counts via tiktoken cl100k_base:
| Tools loaded | Raw JSON | TOON results | Name-only manifest (compact) | Savings vs JSON |
|---|---|---|---|---|
| 5 | 1,519 | 1,003 | 11 | −99.3% |
| 50 | 14,113 | 9,287 | 123 | −99.1% |
| 255 | 71,929 | 47,438 (−34%) | 123 | −99.8% |
Reading it plainly: at 255 tools, raw discovery costs about a 300-page book per session; mcptoon's compact manifest costs a sticky note.
Honest limitations
- Discovery savings apply to tool listing. Per-call arguments and outputs are unchanged — except results encoded as TOON, which measure ~34% smaller.
- Exact numbers vary with your toolset's names/descriptions. Measure your own:
mcptoon demoprints before/after for a live sample server on your machine. - Config-sync CLIs and mcptoon overlap only partially: they move bytes between config files; mcptoon also changes what gets loaded into context and adds a CLI path that works without native MCP support.
Try it
pip install mcptoon # zero deps, ~250KB
mcptoon quickstart # import existing configs, see your numbers
Back to the README · methodology details in tiktoken-benchmarks.md