Agent Comms
A shared source of truth for everyone working on your code — human or agent.
Watch the current terminal tour ·
Animated preview. Recorded from the real TUI using an
isolated demo project; navigation pauses are shortened. No private project
messages or credentials appear in these captures.
Run more than one coding agent on the same project and you already know the failure mode: two of them touch the same file, one grants itself a permission nobody signed off on, and afterward there's no real account of what happened or why. Agent Comms is the layer underneath — leases, signed events, and typed messages that let every agent and every human on your team work the same codebase without stepping on each other. Any agent that can run a shell command or speak MCP joins at the protocol level; Claude Code, Codex, and OpenCode additionally ship a fully autonomous integration out of the box — spawned, prompted, and reported on for you, no wiring required.
[!TIP]
No account, no cloud dependency, nothing to configure before this works. The command block below is the entire setup for a single-developer project.
curl -fsSL "https://raw.githubusercontent.com/DhanushSantosh/AgentComms/v1.1.0/install.sh" | AGENT_COMMS_VERSION=v1.1.0 sh
agent-comms init
agent-comms tui
That's a working project. No server, no config file, no account — a per-project daemon and a local SQLite database start on the first command. (agc is a shorter synonym for agent-comms if you'd rather type that.)
Every action is a signed event, not a log line
Every task claim, message, and approval is cryptographically signed by the actor that produced it — human or agent — and chained into an append-only history. Not a log that can be edited after the fact: a record you can hand to someone else and they can verify it themselves.
Work leases, not hope
An agent claiming src/api locks it for the duration of the work, with a real progress-bearing renewal required to keep it — a heartbeat alone doesn't count. Two agents can't silently overwrite the same code path; the second one finds out before it happens, not after.
The dangerous stuff needs a human
Deleting an identity, granting orchestrator authority, touching credentials or production data — each of these requires an explicit human approval, gated behind a second passphrase-protected signing key that no agent process can reach on its own. This is enforced at the protocol level, not left to a system prompt.
Delivery you can inspect
Agent Comms records delivery evidence separately from an agent's claim and completion. For a live interactive session, it types the request into that session and checks bounded terminal echo evidence before sending Enter. Interactive PTY checks use echo heuristics; signed delivery records attest the recorded outcome.
| Agent Comms | Claude Agent Teams | AutoGen / CrewAI / LangGraph | Jira + Rovo / Monday.com |
|---|
| Works across agent vendors | Any CLI or MCP agent — Claude, Codex, OpenCode ship fully autonomous | Claude Code only | Depends on your own app | Any, via integrations |
| Signed, tamper-evident history | Yes | Not documented | Not documented | Enterprise logs, not agent-signed |
| Approval gates enforced by the system | Yes | Not documented | Bolt-on, custom code | Human task approvals, not agent-action gates |
| Auditable live delivery | Signed records and terminal echo evidence | Internal mailbox | In-process message passing | N/A |
| Setup | One command, local, free | Requires Claude Code | Self-hosted framework | Paid cloud SaaS |
Academic work on the wire protocols agents actually speak — MCP, A2A, ACP — has found they're explicitly not designed to express authorization, audit, or approval workflows. That's the gap this sits in: not another framework for building agents, not another place to run them, but the accountability layer underneath whichever ones you already use.
[!NOTE]
This table reflects a read of publicly documented behavior as of mid-2026, not hands-on testing of every product listed.
Personal mode, the default, is one user and one machine: zero setup, no PostgreSQL, no Docker. Team mode adds a shared PostgreSQL authority when you actually need multi-host coordination — see the service deployment guide.
- Leases last four hours and require real, progress-bearing renewal.
- Shared writes, takeovers, and scope changes require orchestrator-level governance; granting orchestrator itself requires a separate, explicitly human-approved decision.
- Completed work stays active for seven days, then archives without deleting history.
- No telemetry, ever. Update checks are explicit and opt-in.
agent-comms tui for the full terminal control room, agent-comms mcp for a stdio MCP server, readable summaries and tables by default, --json / --output jsonl for automation, and agent-comms doctor for a health check that names exactly what's wrong.
Releases are signed and verified with SHA-256 and Sigstore — see release verification. For the full walkthrough, including wiring up a real Claude Code, Codex, or OpenCode agent as a live participant, start at getting started.
Contributing, or want to run what's on dev before it's released? Build from source — see Contributing. Source builds are unsigned and are for development, not a substitute for release verification.
Documentation · Agent integration · Architecture · Governance · Threat model · Development workflow · Contributing · Release process · Changelog
go test ./...
go test -race ./...
go vet ./...
Worker runtimes speak the open Agent Client Protocol, originally published by Zed Industries — see CREDITS.md.
Licensed under Apache-2.0.
