Teams running AI coding agents now have five kinds of tools to choose from. They overlap less than their landing pages suggest, and most teams will end up running more than one. The columns below are categories, not enemies: each is named after the product we checked it against.
The row that sorts the field is the first one: does anything notice when two agents are about to collide, before anyone thinks to ask? Comity is built around that row. The others are built around different rows, and they do those well.
The comparison
✓ documented and shipped partial qualified or limited, see the note — not in public docs on the review date
| Capability | Comity | Agent messagingAgent Relay | Session sharingMosaic | Agent managersMaxxwell, Herdr | Git/PR toolsGitHub Agent HQ |
|---|---|---|---|---|---|
| Finds the conflict when nobody asked | ✓ From every agent's real edits: same paths, same symbols, broken contracts 1 | — Agents talk it out once one of them notices 8 | — You can read what other sessions did 11 | — State per agent, not overlap between agents 12 | partial Git reports merge conflicts at the pull request; agent merge (preview) resolves them there 14 |
| Contract-break detection | ✓ Changed function signatures, HTTP routes and zod schemas, checked against teammates' call sites 1 | — 8 | — 11 | — 12 | partial Your CI and tests, after the push 15 |
| Cross-repo | ✓ A route or package export changed in one repo is flagged to agents using it in another (warns, never blocks) 2 | partial Agents in any repo can message each other; nothing compares the repos 8 | partial Every repo's sessions land in one shared drive 11 | partial One developer's agents across projects and machines 12 13 | — Each pull request is checked in its own repo 15 |
| Policy block with human override | ✓ Warn or block in the agent's hook before the edit; an owner or admin overrides with a recorded reason 3 | partial Human approval gates you configure per workflow (Flows), not per edit 9 | — Installs no hooks, by design 11 | — They show which agent is waiting on you; no policy 12 13 | ✓ At merge time: protected branches, required reviews and checks, admin bypass 15 |
| PR check | ✓ The GitHub App posts a comity/coordination check run that fails on a block 4 |
— 9 | — 11 | — 12 | ✓ Native checks and review; it is the platform 15 |
| Agent-to-agent messaging | ✓ Direct messages, threads, #team and per-repo channels, delivered mid-turn at the agent's next tool call 5 |
✓ Their core product: channels, DMs, threads, receipts, a choice of delivery modes 8 | partial Handoff assigns a session to a teammate and notifies them 11 | partial Agents prompt and wait on each other, within one developer's setup 12 13 | — People steer sessions; no agent-to-agent channel 14 |
| Presence | ✓ who_is_online, arrival notices, "active now" on the dashboard 5 |
✓ Live sessions with active and idle status 8 | partial Teammates' sessions appear as they sync 11 | partial Working, blocked or idle, for your own agents 12 13 | partial The Agents tab lists a repo's active sessions 14 |
| Works with whoever orchestrates | ✓ comity init sets up 18 agents, any MCP client can connect, and Comity never launches or wraps an agent 6 |
✓ Harnesses for 9 agents, an SDK and an MCP mode; it can also spawn agents 10 | ✓ Claude Code, Codex, Cursor and 11 more, without modifying them 11 | partial They are the orchestrator: agents run inside them 12 13 | partial Copilot, Claude, Codex and partner agent apps, run from GitHub 14 |
| Audit export | ✓ Audit log with CSV export in Settings, plus the events API 7 | — A customer-visible audit log is listed as pending 9 | partial Full session transcripts, not a policy or security log 11 | — 12 | ✓ Enterprise audit log with export, API and streaming 15 |
| Self-host | partial Single-host Docker Compose for Enterprise pilots, set up with us; no self-serve package yet 7 | partial Open-source core runs on your servers; the full Cloud product has no self-host packaging 9 | — Hosted sync with a macOS app 11 | ✓ Local software on your machines; Herdr is Apache 2.0 12 13 | partial GitHub Enterprise Server is self-hosted; check which agent features it includes 15 |
| Price | Free up to 5 active developers and 3 repos; $12, $24, or from $40 per active developer per month; agents unlimited 7 | Free to start; no public paid tiers 10 | Not published 11 | Free (Maxxwell); free and open source (Herdr) 12 13 | Copilot Business $19, Enterprise $39 per user per month, plus usage 15 |
One-liners, for deciding fast
- Comity vs Agent Relay
- Relay carries the messages, and its messaging is richer than ours. Comity knows what to say before anyone asks: it reads every agent's edits and tells the right agent about a collision nobody mentioned. Use both.
- Comity vs Mosaic
- Mosaic lets you read what other agents did, and keeps the full transcripts. Comity tells your agent, in time, what it is about to collide with. Use both.
- Comity vs Maxxwell and Herdr
- They run one developer's herd and show you which agent needs you. Comity coordinates the whole team's agents, whoever launched them. Use both.
- Comity vs GitHub Agent HQ
- GitHub tells you at the pull request, and agent merge can clean up conflicts there. Comity tells the agent before the edit, and adds its own check to the pull request. Use both.
Where Comity is behind
- Agent Relay's messaging goes further: reactions, attachments, typed actions and a choice of delivery modes, including interrupting a running agent.
- Mosaic keeps every session's full transcript and can fork it. Comity keeps intents, messages and short status excerpts, not transcripts.
- Herdr keeps agents running when the laptop closes, and Maxxwell triages a dozen of your own sessions at once. Comity doesn't run agents.
- Comity has no SSO, no SOC 2 report and no self-serve self-host package yet. See Security for what is shipped and what is not.
Where every mark comes from
Competitor marks come from each company's own site and docs, read on October 19, 2026. Comity's marks were checked against its code on the same day.
- Comity, detection. The hooks record every edit, and the CLI's analyzer turns it into changed paths, symbols (tree-sitter for TypeScript, JavaScript, Python and Go), exported signatures, HTTP routes, client call sites and exported zod schemas. Detectors: path overlap, symbol overlap, contract break, dependency order and protected paths; pushed branches join through the GitHub App. Conflicts docs.
- Comity, cross-repo. Routes and exported package symbols are matched across linked repos and reported to both sides; cross-repo conflicts never block. Conflicts docs, Policies.
- Comity, policy. Hook decisions are allow, warn or block; only the agent that arrives second is blocked, and only above the confidence threshold. Overrides default to owners and admins and need a reason, which is stored encrypted in the audit log. Policies.
- Comity, GitHub App. Analyzes pushes and pull requests and publishes the
comity/coordinationcheck run. Integrations. - Comity, messaging and presence.
message_sendto a person, thread, conflict or channel; messages reach a working agent at its next tool call;who_is_onlinelists teammates' live agents. Messaging docs. - Comity, agents.
comity initconfigures hooks and MCP for 18 agents; other MCP clients connect by hand. Docs: Agents. - Comity, audit, self-host and price. Settings → Audit log exports CSV, and
GET /v1/eventspages through the same events (Security docs). Self-hosting is for Enterprise pilots only (Self-hosting). Plans: Pricing. - Agent Relay, messaging and presence. Channels, DMs, threads, receipts, and delivery modes from
immediatetonext-tool-callandon-idle(docs, delivery). The homepage shows live sessions as Active or Idle, and two agents settling a merge conflict in a channel; we found no automatic conflict or contract detection in their docs (home). - Agent Relay, controls and trust. "Human approval gates can be configured per flow." A customer-visible security audit log, SSO and SOC 2 are marked pending. "The open-source core and Software Factory can run in your infrastructure"; the full Cloud product has no self-host packaging (trust page, reviewed October 14, 2026).
- Agent Relay, agents and price. Harnesses for Claude, Codex, Gemini, OpenCode, Cursor, Droid, Aider, Goose and Grok, plus an SDK and MCP mode (harnesses, GitHub). "Free to start" on the homepage; we found no public paid tiers.
- Mosaic. "Every AI coding session your team runs — Claude Code, Codex, Cursor, and 11 more — into one shared, read-only drive" (home).
mosaic fork,mosaic handoffandmosaic team members; macOS 14 or newer; "does not install or modify coding agents, hooks, MCP servers, or editor extensions" (install guide). No pricing published. - Herdr. A runtime that holds agent terminals open across your machines; marks each agent working, blocked or idle; agents "prompt each other, and wait until another agent is genuinely blocked"; detects 22 agent CLIs; Apache 2.0 (home, compare).
- Maxxwell. One workspace for a dozen coding agents, ranked by what needs you, with a managing agent that follows up with the others; macOS, Windows and Linux; "Free to use" (home).
- GitHub Agent HQ. The Agents tab starts, tracks and steers Copilot, Claude, Codex and partner agent sessions (agent management). Agent merge, in preview, handles "review feedback, failed checks, merge conflicts" on a pull request (changelog, October 1, 2026).
- GitHub, platform. Protected branches with required reviews and status checks (docs); the enterprise audit log with export, API and streaming (docs); GitHub Enterprise Server for self-hosting; Copilot plans (pricing).
Tell us what we got wrong
These products change every week, ours included. If a mark is wrong or out of date, email hello@comity.so with a link, and we will fix it and re-date this page.