Slackgentic Team
unbounded-ai/slackgentic-team
Central Slack para coordenar agentes locais Codex e Claude
Ver no GitHubAgent coordenador que transforma um canal Slack em central de controle para agentes locais Codex e Claude. Executa tarefas de código mantendo credenciais e transcrições locais, permitindo orquestração de múltiplos agentes diretamente do Slack sem exposição de chaves de API.
Instalação
npx convez-ai unbounded-ai/slackgentic-team1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Orquestrar múltiplos agentes de código a partir de Slack
- Executar tarefas de desenvolvimento mantendo segurança local
- Revisar trabalho de agentes em paralelo com aprovações em thread
Veredito

Ferramenta madura e bem arquitetada para times que precisam controlar agentes locais sem expor infraestrutura. Excelente integração Slack com design de segurança por padrão, mas requer Python 3.11+ e Codex/Claude já instalados na máquina.
README
A tradução para o português deste README ainda está na fila.
slackgentic-team
Local Codex and Claude agents, coordinated from Slack.
Slackgentic turns one Slack channel into a control room for local coding agents: start work with a normal message, watch agent output stream back into a thread, track sessions started outside Slack, and keep all credentials and transcripts strictly local.
Use agents exactly the way you already do. They still run on your laptop, with the same tools, deployment access, repo access, and approval levels. Slack just becomes the place where you can step away, touch some grass, and reply when an agent needs you.
Why It Exists
- One Slack channel becomes the surface for local agent work.
- Everything stays on your machine.
- The agent channel is private by default, so only you see what your agents are up to.
- Codex and Claude work under one roster.
- No API-key billing detour: it drives the Codex and Claude Code you already use.
- Cross-model reviews are easy: ask for one directly, or let an agent pull in another agent when it wants a second opinion.
- Thread subtasks are natural: reply with
somebody ..., another agent picks up that slice, then the original agent resumes with the added context. - Sessions started outside Slack still get tracked in Slack threads.
- Hire/fire keeps parallelism intentional.
Quick Install
Recommended — installs slackgentic on PATH globally via
uv so the command works in any shell:
uv tool install --with pip git+https://github.com/unbounded-ai/slackgentic-team.git
slackgentic slack setup
slackgentic service install && slackgentic service status
--with pip lets the in-app updater apply new releases in place.
Alternatively, install from a source checkout — needed if you want to track
main or hack on the code:
git clone https://github.com/unbounded-ai/slackgentic-team.git && cd slackgentic-team
python3.13 -m venv .venv && source .venv/bin/activate && pip install -e . && slackgentic slack setup
slackgentic service install && slackgentic service status
Then the CLI will ask you to message yourself in Slack something like:
/slackgentic-<you> setup
Proceed with the rest of the CLI instructions.
Requirements: Python 3.11+, permission to create a Slack app, and codex or
claude on PATH. On macOS, install managed Slackgentic services from Python
3.11-3.13 so Homebrew python3 drift cannot put long-running agents behind a
new TCC prompt identity.
Setup Notes
slackgentic slack setup creates a teammate-specific Slack app, for example
/slackgentic-riley. It uses Slack's official CLI for app creation, opens
Slack-hosted token pages, then saves credentials in
~/.slackgentic-team/config.json with mode 0600.
Choose a suffix explicitly when needed:
slackgentic slack setup --instance riley
Slack Usage
The setup modal creates a private agent channel by default, posts a short command guide, then starts with one Codex agent and one Claude agent. The roster stays updated in-place and pinned when Slack grants the bundled pinning scope.
Run commands with the slash command:
/slackgentic-<you> <command>
Inside the agent channel, you can also type the command body directly, for
example status, show roster, or hire 3 agents.
Useful commands:
/slackgentic-<you> status
/slackgentic-<you> show roster
/slackgentic-<you> show repo root
/slackgentic-<you> repo root ~/code
/slackgentic-<you> hire 3 agents
/slackgentic-<you> hire 1 claude agent
/slackgentic-<you> fire @riley
/slackgentic-<you> fire everyone
status
show roster
hire 3 agents
Start work by typing in the agent channel:
Inspect the repo and summarize the test command
@riley update the README with install steps
Somebody review @riley's PR https://github.com/org/repo/pull/42
#dangerous-mode @riley repair the local service installer
Slackgentic reacts right away, replies in your thread, launches the selected provider, and streams the visible result back. Thread replies continue from the same Slack context.
The roster also has Assign Work and Schedule Work buttons. They open a
Slack form for a task, work/review kind, timing, dangerous mode, repeat cadence,
and an optional dependency on a currently busy agent finishing its active task or
external session.
In the main channel, write anything to hand it to an available agent. Use
@agentname ... when you want a specific agent. Inside a task thread, reply
with somebody ... to bring in another agent for a subtask; the original agent
then picks the thread back up with that new context.
By default tasks launch in safe-auto permission mode: Codex runs with
--sandbox workspace-write and Claude runs with --permission-mode acceptEdits
plus an allowlist for read-only inspection commands such as git status,
git log, git diff, gh pr view, rg, and ls. That keeps reviews and
investigations unblocked without auto-approving anything destructive. Safe-auto
Codex and Claude tasks also receive the configured repo root as an additional
directory so Git worktrees under that root can update their shared metadata
during normal edits and commits.
Add #dangerous-mode to a task when you want that managed agent process to run
with Codex --dangerously-bypass-approvals-and-sandbox or Claude
--dangerously-skip-permissions. Slackgentic strips the tag from the task
prompt and marks active dangerous-mode tasks on the roster.
Thread = Context Boundary
A Slack thread is the continuity boundary. If the same agent continues in that thread, Slackgentic resumes the same provider session whenever Codex or Claude has exposed a session id. The thread transcript is also included as backup context, so handoffs and follow-ups stay grounded in what happened there.
Sessions started outside Slack keep their original Codex or Claude Code session too. Replies from Slack are sent into that live session when possible, or resumed against that exact session id.
Setup asks for a repos root. Agents launch there by default, and phrases like
in sample-app ... can select a sibling repo under that root.
Agents can also talk to each other. When one is unsure, it can ask another agent for a review in the same thread; the router prefers a different provider when a cross-model review is useful.
Agents can schedule one-off follow-ups in the same thread. Slackgentic stores those timers in the daemon state database, so the wakeup is not tied to a provider process or an open terminal session.
PM Initiatives
For larger projects, hand the team a project description and let a PM-flavored agent break it into a DAG of subtasks. There are two entry points:
pm: ship the new logging stack
pm plan the migration to FastAPI
@alice plan the storage refactor
The first two use the pm: / pm <verb> shortcut and route to whichever
PM-kind agent is hired (or fall back to a worker if none are). The third
addresses a specific PM-kind agent by handle — tagging a PM agent always
activates PM mode, even without the pm keyword. Hire a PM with
team hire --kind pm (use --provider claude or --provider codex to
pick the model).
A PM-kind agent owns its initiative end-to-end: it can ask one or two
clarifying questions in the thread before committing to a plan, answers
status questions in the same thread (every reply receives a fresh
[PM HARNESS: initiative state] snapshot of the subtask DAG), and stays
assigned for the life of the initiative. PM-kind agents are reserved
for PM duties — the worker router never picks them up for regular
subtasks.
The PM emits a structured plan (title, summary, up to 20 subtasks with
sibling dependencies). Subtasks may optionally include a co_designers
list of two or more handles, in which case the PM fans the subtask out
into one draft per co-designer running in parallel and follows up with
an automatic synthesis stage. The plan is parked for approval — the
approval message includes a rough cost estimate (subtask count, critical
path depth, dangerous-mode count, and a wall-clock band) so a user can
trim before approving. No subtask runs until the user clicks Start
executing on the plan message (or Cancel to drop it). Once approved,
each subtask becomes a deferred-work row, dispatched the moment its
dependencies satisfy. A watchdog polls active initiatives and surfaces
blockers in the initiative thread — stalled approvals, stuck tasks, or
plan failures — but never starts or stops work on its own. When every
subtask finishes, the watchdog posts a recap and closes the initiative.
At any time during a live initiative, drop pm status (or pm plan,
status, dag) into the initiative thread to get an instant DAG render
without going through the PM agent. The reply lists each subtask with
its current deferred-work status, the worker that owns it, and its
dependencies. The PM agent also receives the same snapshot as a
[PM HARNESS: initiative state] prefix on every follow-up.
Need to course-correct after a subtask fails? Reply with
pm replan: <new context> in the initiative thread (PM-owned initiatives
only). The PM agent receives the failure surface and the remaining
subtask state, re-plans from there, and parks a fresh approval gate.
Marking the PM thread done cancels every non-terminal subtask in the initiative.
Users can schedule future work directly from Slack too:
schedule @riley to check CI tomorrow at 9am PT
schedule somebody review the nightly report every day at 5pm ET
schedule @riley inspect the deploy every Monday at 10:30am America/New_York
schedule somebody check the deploy every 2 hours
schedule @riley check the patio lights during tomorrow's sunset time in Waco
Slackgentic asks an agent to resolve the human-language schedule into a validated hidden control message, so richer requests can use the model instead of a local regex parser. The validated schedule is stored in SQLite and claimed by the daemon when due. One-off items run once; recurring items are rescheduled after each due occurrence.
Task threads have two controls: Finish and free up this agent ends the task
and frees the agent; replying in the thread sends the follow-up into the active
task, interrupting the current managed run first when live delivery is not
available. Replying stop sends only the interrupt without closing the task
thread or replaying fallback-queued replies. Reacting to an agent-authored
message in an active task thread is delivered back to that agent as lightweight
feedback.
Why Hire/Fire Agents?
Different levels of parallelism work for different people. Hire/fire lets you operate at the level that is optimal for you while staying disciplined about not spreading yourself too thin.
If you do not know your level yet, start with the default team of 2. Slowly hire agents until it feels like too many, then back down. That is your sweet spot.
The hard limit is 500 active agents. You definitely do not need that many agents.
Sessions Started Outside Slack
Codex:
slackgentic codex-mcp --install
codex --remote ws://127.0.0.1:47684
Claude:
slackgentic claude-channel --install
claude --dangerously-load-development-channels server:slackgentic
slackgentic slack setup tries to register the Codex MCP server and Claude
channel automatically when those CLIs are on PATH. Run
slackgentic codex-mcp --install or slackgentic claude-channel --install
manually if setup could not reach a CLI, and restart any already-open sessions
after installing. The Claude installer adds Slackgentic's Slack request tools to
Claude's allowlist so they do not require their own approval. No extra MCP flag
is needed after registration unless you start Claude with --strict-mcp-config
or another custom MCP config that excludes user-level servers.
Those commands create tracked Slack threads for sessions you started outside Slack. If all matching team seats are occupied, Slackgentic posts a provider-specific hire button and waits without advancing the transcript cursor, so visible output is backfilled after you add capacity.
If you revive an ended session from its Slack thread, Slackgentic uses a free
matching Codex or Claude seat to resume that exact session. If no matching seat
is free, hire or free that provider; somebody ... can still start a new
session in the same thread using the Slack context.
Service
Install the background services:
slackgentic service install
Common service commands:
slackgentic service start
slackgentic service restart
slackgentic service status
slackgentic service uninstall
On macOS, active managed sessions and sessions started outside Slack keep the
machine awake with the built-in caffeinate command until the work is done.
slackgentic slack doctor also reports macOS network-wake readiness: Wake for
network access, TCP keepalive, Wi-Fi wake support, and scheduled wakes. Network
wake can help with Wake-on-LAN/Bonjour-style wakeups, but it is best-effort and
does not make Slack a reliable way to wake a closed-lid or deep-sleeping laptop.
Updating
You do not need to git pull or rerun pip install to update Slackgentic.
The running daemon checks GitHub Releases of unbounded-ai/slackgentic-team
every five minutes. When a newer published version appears it posts one
prompt in the agent channel:
Slackgentic update available Current:
0.1.0Latest:0.1.1[Upgrade now] [Skip]
Clicking Upgrade now fetches and checks out the release tag in the local source checkout, reinstalls the editable package, restarts the service, and edits the same Slack message in place with a green checkmark once the new daemon is back up. Non-source installs upgrade from the release tarball.
Dirty source checkouts are left untouched until you commit, stash, or move the local changes — Slackgentic refuses to overwrite work in progress.
Reset
Archive the current Slack channel and clear the local pointer:
slackgentic slack close-channel --yes
Reset local runtime state while keeping Slack app credentials:
slackgentic slack reset-state --yes
Recreate Slack credentials only when needed:
slackgentic slack setup --force
Configuration
The config file lives at ~/.slackgentic-team/config.json. Environment
variables override stored values.
| Variable | Purpose |
|---|---|
SLACK_BOT_TOKEN | Slack bot Web API token. |
SLACK_APP_TOKEN | Slack Socket Mode app token. |
SLACK_TEAM_ID | Slack workspace team id. |
SLACK_CHANNEL_ID | Default agent channel id. |
SLACKGENTIC_CONFIG_FILE | Alternate config path. |
SLACKGENTIC_STATE_DB | Alternate SQLite state path. |
SLACKGENTIC_HOME | Alternate home directory for transcript discovery. |
SLACKGENTIC_CODEX_BINARY | Codex executable path. |
SLACKGENTIC_CLAUDE_BINARY | Claude executable path. |
SLACKGENTIC_CODEX_APP_SERVER_URL | Codex app-server URL. |
SLACKGENTIC_CODEX_AGENTS | Default initial Codex count. |
SLACKGENTIC_CLAUDE_AGENTS | Default initial Claude count. |
SLACKGENTIC_EXTERNAL_SESSION_ALLOWED_CWD_PREFIXES | Comma- or path-list of cwd prefixes to mirror for sessions started outside Slack, for example /workspace/repos. When set, sessions outside these prefixes are ignored. |
SLACKGENTIC_EXTERNAL_SESSION_IGNORED_CWDS | Comma- or path-list of cwd path segments or patterns to ignore for sessions started outside Slack. |
SLACKGENTIC_EXTERNAL_SESSION_MIRROR_POLL_SECONDS | Seconds between scans for sessions started outside Slack, default 15. |
SLACKGENTIC_AGENT_AVATAR_BASE_URL | Public HTTPS avatar directory, or off. |
SLACKGENTIC_ALLOW_MACOS_TCC_PROTECTED_PATHS | Allow managed tasks to start in macOS privacy-protected locations after you have granted OS access. |
SLACKGENTIC_UPDATE_CHECK_ENABLED | Enable or disable release checks, default true. |
SLACKGENTIC_UPDATE_CHECK_INTERVAL_SECONDS | Seconds between release checks, default 300. |
SLACKGENTIC_UPDATE_REPOSITORY | GitHub repository used for update checks, default unbounded-ai/slackgentic-team. |
Development
python3.13 -m venv .venv && source .venv/bin/activate && pip install -e '.[dev]'
Checks:
PYTHONPATH=src python -m pytest -n auto tests
python -m compileall src tests
ruff check src tests
ruff format --check src tests
Foreground daemon debugging:
slackgentic slack serve
More validation steps are in docs/e2e-slack.md. The design notes are in docs/architecture.md.
Publishing Releases
- Bump the version in
pyproject.tomlandsrc/agent_harness/__init__.py. - Merge the change to
main.
The release workflow runs on every push to main. When it sees a version that
does not yet have a matching vX.Y.Z tag, it creates the tag and a GitHub
Release with auto-generated notes. Running Slackgentic services consume that
release metadata during their next update check.
Current Limits
- Slack still requires clicking token copy buttons for the app-level
xapp-token. - Managed child processes cannot be reattached after daemon restart. Follow-up replies resume the persisted task/session state, while Slackgentic relaunches the provider process.
- PR review execution passes the PR URL to Codex/Claude; it does not cache GitHub diffs itself.
Coming Next
- Stabilization.
- Link threads and wait for one thread to finish before automatically starting a follow-up task.
- More features.
License
MIT. See LICENSE.