Workspace
- Multi-model chat with per-session engine switching
- Streaming over SSE; reconnect and replay without message loss
- Mid-run steering
- Automatic context compaction + /compact
- Session search / pin / share, file attachments
Apache-2.0 · Self-hosted · v0.6
Put agents inside
your governance boundary
The self-hosted workspace for team agents — agents do the work, your team keeps control. Engines are pluggable (faux / pydantic-ai / ACP driving Claude Code, Gemini CLI), and SSO, auditing, and cost attribution are all built into the community edition.
$ make init && cd deploy$ docker compose -f compose.demo.yml up -d --build --wait# Open http://localhost:3000/register — the first-run wizard guides you to demo modeWHY INAGENT
Governance is the foundation of the community edition, not a plugin behind a paywall
| Capability | InAgent (Apache-2.0 community edition) | The category norm |
|---|---|---|
| SSO (OIDC + PKCE) | Built in — org-level config, JIT provisioning, enforced SSO domains | Commonly behind an enterprise paywall |
| Audit logs | append-only storage, admin UI, CSV export | Usually an enterprise tier or a paid plugin |
| Usage / cost attribution | Four levels (org / workspace / user / session) with an aggregate view | Cost reports are typically a paid feature |
| Trace observability | OTel + trace waterfall UI | Trace viewers are typically a paid feature |
| HITL oversight | Native: command approval cards / structured questions / permission requests routed to approval | Mostly bolt-on toolkits or external layers |
| Coding agent engines | ACP drives Claude Code / Gemini CLI inside the governance sandbox | Locked to a homegrown runtime, or no support for external coding agents |
| Secret governance | AES-GCM vault encryption at rest, keys never echoed back, egress placeholder substitution | Plaintext environment variables or basic encryption |
CAPABILITIES
From conversations to governed team work
Pluggable engines
- faux scripted demo engine (no API key)
- pydantic-ai: four tiers plus a fallback chain
- ACP drives Claude Code / Gemini CLI
- Coding agents run inside the governance sandbox
Governance
- Org / workspace role levels and invites
- OIDC SSO (PKCE + JIT + enforced SSO domains)
- append-only audit log + CSV export
- AES-GCM secret vault; keys are never echoed back
- Three-action command policy: deny / confirm / allow
Sandbox
- One persistent container per (user, workspace)
- Non-root, memory / CPU / PID limits
- Data persists across sessions
- File tree + web terminal
Ecosystem
- Skills: SKILL.md compatible + registry sync
- MCP: stdio / HTTP + OAuth + usage auditing
- Artifacts: five versioned types, with preview for every format
- Anonymous share links
Automation
- cron schedules / webhook triggers (HMAC, replay dedup)
- Native GitHub / Stripe signature compatibility
- Standalone runner service; runs survive api restarts
- Results delivered to Feishu; automation = trigger + specialist
Channels & observability
- Feishu / DingTalk bots running as the bound user
- In-IM HITL button approvals and questions
- OTel tracing + trace waterfall
- Usage attribution across four levels (org / ws / user / session)
First run
- First-run wizard guides you through setup or demo mode
- Keyless demo mode (faux engine)
- A productized 15-minute evaluation path
- demo → full: data stays in place
DEPLOY
docker compose up — your data stays yours
$ make init$ cd deploy$ docker compose -f compose.demo.yml up -d --build --wait$ cp ../.env.example .env$ docker compose up -d --build --wait# demo → full: data stays in placeChina-region registry-mirror overlay, offline air-gapped installs, OTel observability — for the full deployment matrix see the ops manual.
FAQ
Frequently asked questions
What is the SSO tax — and why doesn't InAgent have one?
Plenty of teams only discover at procurement that SSO, audit logs, and usage reports are gated behind an expensive enterprise tier — the community calls this the SSO tax. InAgent's commitment: SSO, auditing, cost attribution, and trace viewing are all in the community edition. The project is Apache-2.0 with no enterprise edition and no held-back feature flags — what you deploy is the complete product.
What are the sandbox security boundaries?
Every (user, workspace) pair gets one persistent container: non-root (uid 1000), memory / CPU / PID limits, and the default seccomp profile. Secrets never enter the sandbox. For production, we recommend a docker socket proxy instead of mounting the socket directly; when you need egress control, enable the egress MITM proxy profile — secrets are injected as placeholders and swapped for real values at the network boundary, with auditing.
Can I skip Claude Code, or swap out the engine?
Yes. Engines are pluggable through the RuntimeAdapter SPI: the faux demo engine needs no API key at all; pydantic-ai targets multiple providers (four tiers plus a fallback chain); and the ACP adapter drives real coding agents such as Claude Code / Gemini CLI. Switching happens per session — you are never tied to a homegrown runtime.
Does it support offline (air-gapped) deployment?
Yes. Run deploy/scripts/airgap-export.sh on a networked machine to export a deployment bundle, then docker load on the target machine and bring the stack up offline. For China-region environments there is also a registry-mirror overlay (covering base images and the pnpm registry), so you can build without a proxy.
How far do the internal IM integrations go?
The Feishu / DingTalk bots run as the bound user and support streaming cards, HITL button approvals and question cards, and artifact delivery. Automation results can also be delivered to Feishu. DingTalk uses a long-lived Stream connection, so no public callback endpoint is needed.