A compiler-backed AI coding agent

What is Benzi? What can it do?

Benzi is a compiler-first agentic harness. It can autonomously investigate bugs, modify an entire project, build features, execute tests, and repair failures. The difference? Benzi compiles your codebase into a single, uniform representation first, and then works on every task using that. Quick facts:

Call flow, data flow, control flow.

A body tells you what it does. The edges tell you what it means. Switch lenses:

reverse blast radius
        main
         └─ run_capture
              └─ Sensor.read   ← you are here
                   ├─ calibrate
                   └─ db.insert   (external)

3 proven callers · 1 possible via dispatch

backflow  Sensor.read(sensor_id)
  sensor_id ← req.query.id   @ api.py:41
  sensor_id ← route.params   @ main.py:12
  sensor_id ← "T-42"        origin: fixture

2 call sites · 1 defaulted arg

def tick(self):
  if self.game_over or self.paused: return
  if self.pending:                       # queued turn
     self.direction = pending.pop(0)
  if self._off_board(head) or head in body:
     return self.end_game()          # ← the guard
  ...

Click a node. Feel the map.

Not a mock-up — this is the shape of the real map. Select a symbol and its callers, its callees, the values that move through it, and the tier behind every edge all come back at once.

Every edge is labelled with how much to trust it.

Precision over recall — a coin-flip would undermine the whole map. So nothing is ever guessed silently; it's tagged.

Every edit gated, receipted, reversible.

Benzi runs every write through the same compiler that built the map, then keeps the map in step with the code. Nothing lands unchecked, and any write can be undone.

1. 01Write The agent changes a line span or a whole symbol, and every write is logged as a step.
2. 02Parse-gated The real language parser checks the result first. An edit that breaks the parse reverts itself.
3. 03Receipt A landed edit reports the changed symbol, its callers, how the call graph moved, and how far the lines below shifted.
4. 04Reindex Only the changed files are recompiled, so every later query sees the code as it is now.
5. 05Rollbackoptional Undo the last N writes from a snapshot, files and map together, with no partial states.
6. 06Runtime traceroptional Run it, then ask Benzi what actually happened: the real arguments, the return values, and which dispatch target won.

Persistent, per-repo memory.

The index is always in context; facts Benzi learns are saved as durable memories, so a future session opens already knowing them — the convention, the gotcha, the bug's real origin.

# a memory that pays for itself later
type    gotcha
name    "config-load-order"
body    "env.py must import before settings,"
        "or SECRET_KEY reads empty. Load-bearing."

# → every future session opens knowing this

Every tool answers one question.

Every lookup is O(1): the compiler has already resolved the answers into hashmaps, so the tools read them instead of searching. This list shows 16 of Benzi's 36 tools. Full tool list here.

Measured, not asserted.

One configuration, one attempt per instance, no ensembling, no test-time selection — graded by the official SWE-bench harness.

FAQ & comparisons

Nine questions, straight answers — what runs where, what it costs, and how Benzi differs from the tools you already know.

Two ways in.

Run Benzi's own agentic loop — in the editor or from a terminal — or hand the compiled index to the agent you already use. Bring your own key; free during beta.

$ benzi-login --login you@example.com
$ benzi-headless . "fix the failing test in parser.py"

$ claude mcp add --scope user benzi -- benzi-mcp

$ claude mcp add --scope user benzi -- benzi-mcp /path/to/your/repo