Your agents say done.
The evidence gets a vote.
AgentMagnify turns coding-agent work into a live, replayable project record, and keeps what the agent claims visibly separate from what the tests, builds and screenshots actually support.
$ npx agentmagnify install
Works with Claude Code, Codex, Cursor and any agent that can run bash.
Progress breakdown
5 points of reported progress are not yet backed by evidence.
Needs review
“Checkout flow UI” was reported complete with no verification. Then 3 of 8 e2e tests failed at the payment step.
An illustration of the panel, drawn from its own demo project.
Agent activity is not the same as project progress.
A busy terminal proves nothing. AgentMagnify tracks three numbers and refuses to blend them, because the distance between them is the most useful thing a monitoring tool can tell you.
Reported — what the agents claim
Every task, test and completion the agents send, exactly as they told it.
Roadmap — what the plan implies
Weighted arithmetic over the declared roadmap. Never a feeling, never elapsed time.
Verified — what the evidence supports
Only completions backed by passing tests, builds or reviewed evidence. An unverified “done” stays visibly a claim.
Add visibility without adding another workflow.
No CLI daemon, no IDE plugin, no ceremony. A bash-only skill your agent carries; the panel does the rest.
Pair in seconds
One command pairs the machine, one connects a project. Nothing is watched until you say so, and no secret is ever displayed for copying: not in the panel, not in the terminal.
Agents report as they work
Structured events at real boundaries: tasks, tests, builds, blockers, decisions. Offline, everything queues and the work never stops.
You watch what’s true
A live board of every project, an attention queue of what genuinely waits on you, and a replay of how it all happened.
Claims come with receipts.
Agents upload screenshots and test reports straight to your panel’s storage: a browser test becomes a numbered, pass/fail-badged strip of screenshots under its run. When a completion has no evidence, an independent observer says so, out loud, on the project’s front page.
Screenshot evidence
“The login page works” arrives with a picture of the login page working, attached to the claim it proves.
Test flows you can read
E2E runs render as step-by-step visual stories. The failing step is always there, in red, with the screenshot that shows why.
An observer that disputes
A separate verification pass compares claims against test results. A contradiction is the headline, never a footnote.
Being watched changes the work. Mostly for the better.
A panel on agent work is not neutral: it changes how agents behave and how you spend your attention. Here is what actually moves, in both directions — we measure our own product with the same honesty it demands of agents.
Fewer false “done”s
When a completion only turns green with evidence behind it, declaring victory early stops paying. The gap between reported and verified is on the front page — and visible gaps close. Rework surfaces in minutes, not at review.
Your attention, spent once
Sub-agent questions route to their orchestrator, and only what genuinely needs a human reaches you. Fewer context switches for you; less waiting for the agents. The attention queue is short because it is honest.
The honest costs
Reporting at boundaries costs something: one bash call per task, test or build, and a few hundred tokens of session overhead. A confused agent skips a report rather than blocking — events queue offline, the work never stops, and a gap shows as quiet, never as a fake green.
Replay the project — not just the conversation.
Everything is an append-only event, so the panel can rebuild any moment of the project and play it back like a tape. The rest of the toolset follows from taking that record seriously.
Live Board
Every project in the workspace on one screen, ranked hottest-first: progress bars, warnings, and who is working right now.
Attention Queue
Only what genuinely waits on a human. Mark items as seen without losing them; agent-internal questions never ring your bell.
Replay
Scrub through the whole project as recorded events — progress, blockers and decisions reconstructed at every frame.
Decisions that route themselves
A sub-agent’s question goes to its orchestrator. Only questions that truly need a human reach you, and answered ones say who answered.
Offline-first reporting
Events queue locally, deduplicate on delivery, and replay after an outage. A network failure never blocks the agent’s real work.
Multi-agent aware
Logical agents grouped by the real session they run in, each reporting its own work. The orchestrator is never a bottleneck.
Works where your agents work.
Claude Code
npx agentmagnify install
/plugin marketplace add remanatech/agentmagnify-skill
/plugin install agentmagnify@agentmagnify
One command installs and self-tests it. The marketplace lines are the alternative, with native /plugin update versioning.
Any bash environment
npx agentmagnify install
Codex, Cursor and anything that can run a shell script.
CI & containers
AGENTMAGNIFY_TOKEN=…
A scoped token in the environment; nothing interactive required.
Cloud agent sessions
/agentmagnify BCDF-GHJK-MNPQ
Paste a single-use pair code into ephemeral sessions. Ten minutes, one use.
Start free. Upgrade when a ceiling matters.
Free
Free
Your own machine, your own projects.
- 1 active project
- 1 seat · 1 agent access key
- 14 days of event history
- 256 MB evidence storage
- Live board, replay, attention queue
Pro
$30/year
A person running serious agent work.
- 3 active projects
- 2 seats · 5 agent access keys
- 90 days of event history
- 1 GB evidence storage
- Everything in Free
Business
$150/year
Teams supervising agent fleets.
- 15 active projects
- 10 seats · 30 agent access keys
- 180 days of event history
- 10 GB evidence storage
- Roles, audit log, team groups
Enterprise
Custom
Your infrastructure, your rules.
- No ceilings
- SSO / SAML, SCIM provisioning
- Self-hosted or private deployment
- Custom retention, IP allowlists
- SLA
Prices are quoted in your currency at checkout, inside the panel.
Every plan is a subscription to the service. The numbers above are limits on concurrent use — not credits, tokens or counters, and nothing is spent down as you work.
Project intelligence without handing over your codebase.
Metadata, never code
Events carry task names, counts and commit hashes. Never source, secrets or terminal output. A local filter enforces it before anything leaves the machine.
Secrets stay invisible
Pairing never displays a token for copying, in either direction. Credentials live as digests server-side and 0600 files on the machine.
Self-host it all
The whole platform runs from a docker compose file, evidence storage included. The agent skill is MIT-licensed and open on GitHub.
Magnify your next agent-built project.
Pair one agent and watch the first event land. The free plan needs no card and no configuration.