Build from what happened.
Receive projected events for memory, audit, usage, reports, telemetry, or your own tools.
RECEIVES EVENTS · CANNOT REPLY
Pitot / open coding-agent transport
Add the behavior your coding agent is missing.
Pitot connects ordinary local code to the coding agents you already use.
Observe agent activity, return decisions, and build approvals, verification, memory, audit, or your own workflows—without switching agents or maintaining a different integration for every host.
LINK ACTIVETRACE 0042 · 18 ms
01 / THEIR CODING AGENT
ClaudeCodex
CopilotCursor
GeminiKimi
OpenCodePi
Qwen
02 / PITOT
PitotNORMALIZE
PROJECT EVENT
CORRELATE REQUEST
VALIDATE RESPONSE
PROJECT EVENT
CORRELATE REQUEST
VALIDATE RESPONSE
03 / YOUR CODE
CONSUMER
Observe what happened.
Receives projected agent events.
Cannot affect the waiting action.
Cannot affect the waiting action.
CONTROLLER
Decide what happens next.
Receives a registered request.
Returns one correlated response.
shell.execute → allow
Returns one correlated response.
shell.execute → allow
THEIR AGENT → PITOT → YOUR CODEPITOT TRANSPORTS. YOUR CODE DECIDES.
SCROLL TO OPEN THE BOUNDARY ↓
Boundary / 01
Two roles.
One hard line.
Observe what the agent does, or decide what happens next. Pitot keeps those capabilities separate.
Add behavior at the action boundary.
Receive a registered request and return one correlated decision to the coding agent.
RECEIVES REQUEST · RETURNS ONE RESPONSE
Build / 02
What would you add to your coding agent?
Once the boundary is open, the missing feature becomes ordinary code.
Remember the useful parts.
Build selective project memory from projected events, without preserving entire chats.
Verify before continuing.
Check tests, CI, repository state, or another source of truth before the agent proceeds.
Ask for approval.
Route a release or protected action through your own local approval logic.
Measure what actually happens.
Track actions, model usage, latency, failures, and outcomes across coding-agent hosts.
Build something new.
Use the same boundary to create workflows and tools that coding-agent vendors have not built yet.
Intervention / 03
You do not need to replace the whole agent.
Different tools intervene at different points.
SWITCH AGENTS→
replace the product
ADD A SKILL→
advise the model
ADD AN MCP SERVER→
expose tools and data
USE PITOT→
connect your code to observable events and waiting actions
Quickstart / 04
From an idea
to a live agent hook.
Create a tiny Consumer or Controller in your language. Run it against a real coding agent, edit the code, and immediately change what your tool observes or decides.
01
go install github.com/operatorstack/pitot/cmd/pitot@latest02
pitot init --template shell-policy --language go --dir ./agent-policy03
pitot doctor --host kimi04
pitot dev --host kimi -- kimi -p "Run: echo hello"Supervised adapter proof / 05
Test the whole loop.
Not a mock boundary.
Pitot tests the complete path through real coding-agent hosts:
prompt → hook → projected event → Controller decision → tool result
Not just generated payloads.
Not just protocol unit tests.
The actual integration boundary.
UBUNTU · macOS · WINDOWS
01REAL PROMPT
02HOST HOOK
03PROJECTED CONSUMER
04CONTROLLER ALLOW / DENY
05TOOL RESULT
HONEST SCOPE
Pitot transports the event and the decision. Your code decides what approved, verified, safe, complete, or useful means.