Nº FEATURES + PROOF
Features & Benchmarks
Task-led capabilities, then the hop-count proof: six questions, five fixture repos.
- 01Orient in a new repo
DNA, map, landmarks, and stack profile.
- 02Prepare a change
Pack, reuse, House Style, and a work order — never file bodies.
- 03Blast before you edit
See dependents, tests, and features in the blast radius.
- 04Guard a session
Rank this session’s dirty files. Read the high-blast ones first.
- 05Envelope and Iris
Index facts plus a receipt for every Intelligence call.
- 06Queue a finding
Unused, duplicate, wrapper — resolved only after a re-scan.
- 07Review a change
Diff-aware impact for pull requests.
- 08Delete safely
Check whether a symbol or file still has dependents.
- 09Track health
Engineering health scores and history.
- 10Domain deep dives
Frontend, backend, testing, and security reports.
- 11CI gates
Run Prism in pipelines with fail-on thresholds.
- 12Dispatch a teammate
Hand a change to a background teammate in your editor; Prism runs the checks when it stops.
- 13Start my day
A standup briefing: yesterday's jobs, what's waiting, suggested focus.
- 14Full capability table
Every command and MCP tool in one lookup.
What we measure
The same six questions an agent asks before it edits, on five fixture repos. One side lists, greps, and skims. The other calls Core — the same tools MCP exposes. We count discrete tool calls — the hops an agent takes before it can answer.
Latest run
Recorded Sep 4, 2026, 2:18 PM UTC · darwin/arm64 · v24.3.0
By question
Summed across the five fixtures — list/grep on the left, a Prism tool on the right.
| Question | Without | With Prism | Calls saved |
|---|---|---|---|
| What is this repository? repository_dna | 64 | 15 | −77% |
| Is this edit safe? blast_radius | 58 | 10 | −83% |
| Is this codebase healthy? repository_health | 75 | 15 | −80% |
| Where does this symbol live? find_symbol | 53 | 10 | −81% |
| Which tests should I run? test_impact | 22 | 10 | −55% |
| Are there import cycles? dependency_cycles | 53 | 10 | −81% |
How to read this
The win is fewer round-trips and a complete structural answer. Naive grep grows with file count; Prism answers stay one or two calls. Wall time includes a cold index per question so the runs stay isolated; a real session pays that once.
Re-run locally after bun run build: bun run bench:orientation.
One-click MCP install
Add Prism to your agent in one step, then ask in plain language.
{
"mcpServers": {
"prism": {
"command": "npx",
"args": [
"-y",
"--prefer-online",
"@repo-prism/mcp-server@latest"
],
"env": {
"NODE_USE_SYSTEM_CA": "1"
}
}
}
}Project root, or ~/.cursor/mcp.json for every project. Then Settings → MCP → enable prism (51 tools).