Works with Claude Code and Cursor · Runs locally

The supervision layer for AI coding.

Watchpost runs beside Claude Code and Cursor on every engineer's machine. It flags risky changes as they happen, verifies understanding before code merges, and turns every catch into a guardrail the whole team inherits.

$ npx watchpost init
✓ watching · session: fix-token-refresh

claude codefix-token-refresh

{{ ln.t }}

watchpostwatching

▲14:34 · attention
removed the auth check in verifySession() — contradicts the stated task
[a] ask about this · [d] dismiss

{{ f.time }}{{ f.tier }}{{ f.text }}

CHECKPOINT · TURN END

What happens when the refresh token is expired?

I can answerDraft steering prompt

flags {{ flagCount }}checkpoints {{ checkpointCount }}{{ cost }}

01 — THE PROBLEM

Your codebase is being written faster than anyone reads it.

11.4 hrs/week

time developers now spend reviewing AI code, vs 9.8 writing it

2026 developer survey

5.3× longer

how much longer agentic-AI PRs wait for reviewer pickup — 1,055 vs 201 minutes

LinearB 2026, 8.1M PRs

+242%

rise in incidents per PR as AI code volume grew

Faros AI, 22k developers

The bottleneck moved from writing code to understanding it.

02 — IN EVERY SESSION

Watchpost sits beside the agent, not in its way.

watchpost · session: fix-token-refresh● watching

Live narration

Every change described in plain language as it lands, tiered by how much attention it deserves. Risky edits pin to the top with the reason.

31routineread src/api/client.ts and src/auth/session.ts
32routineedited src/api/client.ts (+41 −7)
33notableadded retry logic to fetchUser()
34▲ attnremoved the auth check in verifySession()
35routineformatted imports in client.ts

Prompt interceptor

Oversized or bundled prompts get caught before submission, with a ready-to-paste plan to run them as parallel tasks instead.

▲this prompt bundles 3 independent tasks
1migrate the sessions tablerun first
2add refresh-token rotationparallel
3update the client retry logicparallel
ppaste split plan↵send anyway

Checkpoints

One targeted question before a change is accepted. If the engineer can't answer it, Watchpost drafts the steering prompt that makes the agent explain or fix.

CHECKPOINT · TURN END
What happens when the refresh token is expired?
aanswerddraft steering prompt

Receipts

One command turns the session into a PR-ready summary: what the agent changed, what the engineer verified, what was flagged and resolved.

$ watchpost receipts
→agent: 3 files, +58 −19
→verified: 2 checkpoints answered, 1 flag resolved
→copied to PR description

03 — ACROSS THE TEAM

One engineer's catch becomes the whole team's guardrail.

Receipts, team-wide

Every AI-assisted PR opens with its receipt: what the agent changed, what the engineer verified, what was flagged and resolved — rendered right in the PR description. Reviewers start from proof instead of archaeology.

Guardrails

Upvoted lessons compile into rules: CLAUDE.md/AGENTS.md entries and pre-push hooks that fire before the same mistake ships twice. Each rule counts its saves — prevented ×7.

The digest and playbook behind them

Engineers approve every lesson before it leaves their machine, edited and anonymized. The digest ships when there's something worth reading — cadence adapts to volume, never padded. Every lesson lands in the searchable playbook: the fastest answer a new hire gets to "how do we use agents here."

Adoption dashboards tell you usage went up. Watchpost changes what happens in the next session.

04 — BUILT FOR SECURITY REVIEW

Supervises agents, not engineers.

Leaves an engineer's machine: lessons they approved · anonymous counts.

Never leaves: prompts · transcripts · diffs · code · filenames · identities.

Session data lives in local SQLite. Model calls run through each engineer's own account. There is no code access, no CI integration, and nothing for your security team to host. Team trends are aggregate-only and hidden below five members. Individual data has no path to a manager — by architecture, not policy.

Read the security page →

05 — ROLLOUT

Pilot in a week, not a quarter.

npx watchpost init

Two minutes per engineer, no admin rights, no code access.

Value on day one

Flags, checkpoints, and receipts work solo, fully local.

Switch on the relay

Digest, playbook, guardrails, and trends for the team.

Flat-priced three-month pilot for one team — see Pricing.

Start free on your own machine. Pilot it with one team.

You're on the list.

06 — QUESTIONS

Questions

{{ f.icon }}