measured, not promised
The numbers, with the method and the mistakes
tessera's whole pitch is a comparison: what your agent reads today vs. what a compact,
cached, metered answer costs. So here are the tables — including the row where we lost.
1 · Whole-file reads vs. atlas cards (reference workload)
Same questions, two ways of answering. Lower is better; tokens are the counting unit.
| Question type | Whole-file read | tessera card | Ratio |
|---|
| Where is this symbol defined? | 22,800–46,450 | ~100 | 1.6–302.8× |
| What calls this function? | 38,410 | ~220 | ~175× |
| Which tests cover it? | 12,900 | ~140 | ~92× |
| What did we decide about X? | thread archaeology | 41 | uncountable |
2 · The row we fixed in public
code_tree returned far more than it displaced — the meter caught it, and it stayed in the report.
| code_tree | Returned | Displaced | Net |
|---|
| Before (uncapped, June) | 367,045 | 52,160 | +314,885 |
| After (paged outlines) | 10,723 | 22,820 | −12,097 |
3 · What a session costs before you ask anything
Tool schemas are context tax. Default matters.
| Profile | Schema tokens | Tools | Switch |
|---|
| core (default) | ~1,200 | 7 | — |
| memory (close-out) | ~900 | fewer | set_profile |
| full | ~2,900–4,627 | 15 | set_profile (mid-session) |
4 · Deployment shapes
The same server, three ways.
| Shape | Best for | Cost | Notes |
|---|
| Local single-user | one dev, one machine | $0 | keychain secrets, local Postgres |
| Self-hosted team stack | 2–50 people | hosting only | tessera stack scaffolds compose + TLS + backups |
| tessera Cloud | teams without ops | from $8/mo | in development — hosted MCP + shared Postgres per team |
Run the atlas yourself (in-browser, 120 KB wasm)
all five directions →