Skip to content
Local-first · provider-neutral · evidence-backed

Command your agents. Inspect their work.

Commandem is the local-first orchestration layer for coding agents. Plan work across projects, route it to the right provider and account, monitor runs, and inspect results with gates and receipts. Commandem can check work when a Mission has a gate; work without a gate remains unverified.

Private builds today; public release in preparation. No Commandem cloud account required. Model providers receive the context you choose to send.

Commandem · Overview

Overview: real window capture of Commandem on real project data.

Real window capture of Commandem on real project data. Cost estimate hidden. Open full-size Overview image.

✓ Native macOS app✓ Claude · Codex · Gemini and more✓ Named accounts and Profiles✓ Local receipts and policy

Command + proof

A complete orchestration loop.

Commandem coordinates agentic work with explicit planning, routing, human control, and deterministic gates. It keeps intent, execution, evidence, and explanation connected so “done” means more than an agent saying it.

01PlanTurn Campaigns into bounded Missions with dependencies and gates.
02DispatchRoute a Profile, account, model, and Loadout explicitly.
03MonitorSee runs, approvals, failures, quota, and stale sources together.
04GateTest the claim against the required change and evidence.
05ReplayReconstruct what happened and keep the receipt.

Plan

The roadmap is an executable workspace.

Search and scope work across projects, schedule capacity, inspect dependencies, and require a gate before dispatch. Campaigns hold the objective; Missions hold the work.

TableBoardTimelineDependencies

Plan

Plan: real window capture of Commandem on real project data.

Plan · Real window capture of Commandem on real project data. Cost estimate hidden. Open full-size Plan image.

Monitor

One causal signal stream for the whole fleet.

Runs, provider capacity, approvals, privacy signals, notifications, and service health share one document. Open a signal to see the observation, gate result, cause confidence, freshness, and canonical source.

SignalsApprovalsCapacityReceipts

Monitor

Monitor synthetic provider signals showing quota exhaustion, a partial-source warning and an inspector.

Current work loop

See the plan, its limits and the recorded result.

Overview, Usage, Launch and Debrief are real window captures of Commandem on real project data. The navigation strip is a component example.

Seven project destinations: Chat, Overview, Plan, Monitor, Review, Code and Terminal.
Overview activity: real window capture of Commandem on real project data.

Overview activity · Real window capture of Commandem on real project data. Cost estimate hidden.

Debrief: real window capture of Commandem on real project data.

Debrief · Real window capture of Commandem on real project data. Cost estimate hidden.

Inspect Usage and Launch examples
Usage: real window capture of Commandem on real project data.

Usage · Real window capture of Commandem on real project data. Cost estimate hidden.

Launch: real window capture of Commandem on real project data.

Launch · Real window capture of Commandem on real project data. Cost estimate hidden.

Evidence, not vibes

Experiments measure effects. Replay explains runs.

Commandem separates descriptive reconstruction from causal claims. Experiments own measured comparisons; Replay shows the trace and labels what remains unknown.

Experiments

Experiments index showing declared evaluation suites, status, cell count, and spend.

Experiments: every declared study and its status, so a negative or pending result stays visible, not hidden. Sample data

Replay

Replay workspace with a timestamped trace, causal inspector, alternatives, and explicit claim grade.

Replay: the recorded trace behind a run, with what's observed kept separate from what's inferred. Sample data

4,370experiment runs indexed across the current research archive; coverage and comparability vary by study.
95–0favorable-to-harmful discordant matched impossible tasks in the audited study; runner versions differed.
25rejected completion claims recorded in sampled product telemetry, rather than silently converted into success.

Frozen audit snapshot, September 2026. Counts establish observed behavior in the recorded runs; they do not establish a universal model effect.

Read the dated research context and limitations → These figures describe different populations and must not be pooled into a single reliability rate.

Context you can inspect

Know what each turn received and what it did not.

The Context inspector records source identity, precedence, permission, retrieval, exclusions, retention, redaction, and truncation. Profiles bind account and project behavior without hiding the effective configuration.

Knowledge + context

A manifest before memory becomes magic.

Cross-project recall and Matrix recommendations stay bounded by source and privacy policy. Context can be reviewed, removed, and forked into an experiment instead of becoming invisible prompt state.

Source identityRedactionRetentionExclusions

Context

Context inspector showing delivered, considered, and available sources with provenance and retention detail.

Profiles + privacy

Make the effective account and data posture visible.

Profiles collect project references, account bindings, Loadouts, models, tools, notification modes, and privacy posture. Named choices stay named in the run and receipt.

Multi-account routing

Bind an account per role and make failover explicit. No silent switch when quota runs out.

Privacy gates

Review what can leave the machine before dispatch; unknown provider settings remain unknown.

Durable receipts

Record account, model, policy, gate, evidence, and outcome for later review.

Notification modes

Prioritize, silence, or schedule alert types without discarding their history.

Local-first, precisely stated

Commandem does not require a hosted Commandem account. Its registry, policy, history index, and receipts are local.

  • Your selected model provider receives the prompts and context you send.
  • Optional GitHub and sync adapters contact their services.
  • The local safety floor is a guardrail, not an OS security boundary.

Read Privacy →
Read Security →

Extend the fleet

Reusable capabilities with distinct trust boundaries.

Recipes render prompts, plugins execute code with declared permissions, and skills give agents specialized instructions. The marketplace keeps those mechanisms separate while making them searchable together.

Public release path

Private builds today. Public release in preparation.

The current build is actively tested with real multi-agent work. Key rotation, final release gates, packaging, and the public repository decision remain before launch.