Termada

Termada

Persistent, policy-controlled missions for AI agents operating real environments.

[Release] [CI] [MCP Registry] [Go] [License]

Termada is a single-binary, local-first runtime that sits between an AI agent and the terminal, locally or over SSH. The agent talks to it over the Model Context Protocol and gets a sturdy toolset instead of a raw shell: bounded exec_run waits, persistent sessions that keep cwd/env, async jobs with streamed output, PTY input for interactive prompts, and structured results — while you watch and control active jobs from a live dashboard with a job kill-switch and an approval queue. Mission Control adds a durable goal, GPT/Codex-authored plan, runtime-verified steps, interruption recovery, and an exportable evidence report around that execution.

[Termada — the live dashboard]

_{The live dashboard: every session is a real terminal you can watch and take over — block or pause the agent, type in yourself — beside the agent panel, policy management, a tamper-evident History, and a Stop-All active-job kill-switch.}

Why

Handing an AI agent a raw shell is fragile and opaque: a command blocks on a prompt and the agent hangs; cd and exported env vanish between calls; long builds flood the context window; and you can't see — let alone stop — what's running. Termada replaces the raw shell with a runtime that is reliable for the agent and transparent for you:

• Reliable for the agent — blocking exec_run calls use bounded wait budgets and return structured output; sessions persist cwd/env; long jobs run async and stream incrementally instead of dumping; interactive prompts are answerable.
• Transparent for you — one dashboard shows every agent and every session as a real terminal; controlled actions matched by confirm wait for your approval; one button installs a short-lived action admission barrier, requests cancellation of active gateway executions, signals active engine jobs and closes forwards; remote/provider effects still require terminal verification or reconciliation.
• Accountable as a mission — an agent's plan, real commands, approval decisions, verification jobs, outcome, and audit anchors stay together. A plan step cannot pass without a real mission job that Termada observed exiting with code zero.

Mission Control

For a concrete operational outcome, Codex or another MCP agent starts with mission_create(goal, plan). Termada allocates a dedicated persistent session; normal exec_* calls keep using the same policy, approval, PTY/SSH and audit paths. The agent attaches each successful job_id with mission_update, then exports mission_report.

[Termada Mission Control waiting for a human approval]

The report separates runtime-observed evidence from agent notes, includes exact audit sequence/hash anchors, and returns a SHA-256 recorded in the audit. If the daemon restarts, a non-terminal mission becomes interrupted; mission_resume creates a fresh session attempt instead of pretending PTY state survived.

Run the isolated real-execution demo:

./demo/mission-control/run.sh

It starts a broken loopback HTTP service and a separate demo daemon. The protected repair genuinely changes service state and waits for a human approval; the final probe must observe HTTP 200. See demo/mission-control and the 2:40 demo script.

Features

Execution engine

• Persistent-shell sessions over a PTY that keep cwd, env, and venv between commands.
• Async jobs: exec_start → job_id; poll incrementally by a stable cursor, with sequential bounded pages, a full status state machine and structured errors.
• Answer interactive prompts (exec_write, with secret redaction). Local PTY jobs support process-group signals/kill; remote SSH interrupt/kill requests are best-effort Ctrl-C, not a guaranteed force-kill.
• Clean output: stateful ANSI/VT stripping, CR-collapse, bounded retention, best-effort secret redaction.

Live control & observability

• A long-lived daemon with a control plane over a Unix socket; serve --stdio is a thin shim that proxies MCP to it — so multiple agents share one daemon and one dashboard.
• Web dashboard where each session renders as a real terminal (xterm.js, streamed over SSE) with operator take-over: type into a job's PTY, hold the agent's input, or pause its output.
• Workspace labels, bounded state bootstrap, cursor-resumable live updates, English controls and responsive desktop/mobile navigation.
• Approval queue, activity feed, policy/server management, and a Stop-All cancellation barrier for admitted actions, active engine jobs and forwards.
• A TUI (termada top) and a full inspection CLI.
• Synchronously recorded, hash-chained, best-effort-redacted audit log. termada audit verify verifies the continuous chain across rotated segments.

Mission Control

• Durable, owner-scoped missions with a goal, up to 24 plan steps, dedicated session, multiple resume attempts, bounded evidence timeline and terminal outcome.
• Runtime enforcement for passed steps (job_id from a mission session, status=exited, exit_code=0) and refusal to succeed while jobs remain active.
• Mission-first desktop/mobile dashboard with pinned approval, stale/offline states, compact plan/evidence views and Markdown report download.

Security

• Policy engine: every argv word is shell-quoted; allow / deny / confirm matching sees leading assignments, absolute paths, Darwin case variants, known wrappers and explicit shell payloads. Shell scripts/stdin/interactive shells and ambiguous compound commands fail closed when deny/confirm rules exist. The structured policy additionally matches authenticated principal, tenant, typed action/resource/context, risk and reversibility. Every controlled action that resolves to confirm parks in the durable operator queue, times out to deny, and can execute only when the caller retries the exact approved action id.
• One canonical gateway covers every public side-effect route: current-policy recheck, one-use approval consume, atomic budget authorization, durable intent, provider outcome and explicit outcome_unknown reconciliation state.
• Policy-owned budgets cover operations, runtime, monetary amount, tool/model cost, production changes and post-outcome failures at exact global, tenant, principal or recipient scopes.
• age-encrypted vault (no CGO); vault APIs never return secret values to agents. Values are injected daemon-side and registered for best-effort output redaction.
• Per-agent quotas and owner isolation for jobs, sessions, session-scoped remote file operations and forwards. Sessions and pending confirmations are capped at 32 per owner and 128 total; live forwards at 16 per owner and 64 total. Local host paths remain a shared OS/filesystem boundary. Set security.identity_mode: strict in production: every configured agent must have an existing assigned policy and authenticate with either a unique strong static token or an explicitly enabled, offline-verified short-lived agent_identity JWT. Tokenless, unknown and policyless callers fail closed. A strict policy without structured rules also denies every command that lacks an explicit legacy allow match; deny/confirm gates take precedence. development retains self-asserted ids for trusted local clients and is the compatibility default for older configs that omit the field.
• Optional security.run_as drops local shell processes to a dedicated uid and disables daemon-privileged local file tools and daemon-environment inheritance; use commands inside that dropped session for local file access. Remote SFTP remains available.
• Every TCP /api/* request and /metrics requires either the dashboard token or an enabled, pinned-JWKS-verified human JWT. Operator-only Unix-socket routes similarly require either a verified human JWT or the separate cli.token used by the local CLI.
• Optional pinned-JWKS operator identity binds subject, tenant, roles, delegation and verified MFA to approvals; policy can require role quorum, distinct approvers and requester exclusion. Optional Ed25519 audit anchoring exports signed fsynced chain heads through a durable HTTPS outbox.

Remote & fleet

• Persistent remote SSH sessions with reconnect — a dropped link is re-dialled as a fresh shell so the session can serve new commands. The authorized name, host, port, user, tags and keyed vault-reference binding stay pinned, so inventory edits cannot retarget an approved or live session. Prior cwd/env are lost; an in-flight job becomes orphaned, and its uncontrolled remote process may still continue. Verify remote state before retrying it.
• fleet_run across servers by name or tag with best-effort-redacted, structured per-server results. Commands must be non-empty argv arrays and run under a shared daemon-wide ceiling of five concurrent fleet targets; a call may request less. One call matches at most 256 targets and returns at most 2 MiB of aggregate result text. SSH uses vault creds, ssh-agent, or on-disk keys, with serialized, fsynced TOFU host-key pinning that fails closed on a malformed known_hosts file.
• Owner-scoped local-to-remote port forwards. Opening one is policy-gated and rolled back unless its start audit record is durable; listeners are loopback-only and bounded to 64 simultaneous connections per forward.

Operations

• Crash recovery (jobs persist; running jobs come back as orphaned), bounded local-FS snapshots/undo, desktop notifications and outbound-only Telegram notifications.
• A typed adapter SDK provides durable prepare → execute → verify → compensate/reconcile phases, provider idempotency keys, secret-free receipts and restart recovery. Generic shell and plugin effects remain opaque unless a concrete provider adapter defines and proves its semantic invariants.
• Out-of-process plugins exposed to agents as <plugin>.<tool>. Plugins are trusted executables, not a security sandbox; calls are policy-gated and do not start unless their start audit record is durable.
• termada update — bounded self-update from GitHub releases on Unix (mandatory SHA-256 verification, optional Ed25519-signed checksums, exact-member extraction, atomic replace). Windows reports an explicit manual-install path because a running .exe cannot be replaced atomically.

Not yet: a native Windows ConPTY runtime (cross-compiles today, but PTY and signals are stubs) and code-signing / notarization.

Install

One line, no Go needed — downloads a prebuilt macOS or Linux amd64/arm64 binary (SHA-256 verified) to ~/.local/bin:

curl -fsSL https://raw.githubusercontent.com/Islomzoda/termada/main/install.sh | sh

Pin a version with TERMADA_VERSION=vX.Y.Z, or change the location with TERMADA_BIN_DIR=~/bin. If ~/.local/bin isn't on your PATH, the installer prints the one line to add.

# Docker (current published image: linux/amd64). Keep the published port on# host loopback; the image binds 0.0.0.0 only inside its network namespace:docker run --rm --platform linux/amd64 -p 127.0.0.1:7717:7717 ghcr.io/islomzoda/termada
# Persistent container state runs as uid/gid 10001. A fresh named volume is# initialized with the image's ownership:docker run --rm --platform linux/amd64 -p 127.0.0.1:7717:7717 \  --mount type=volume,src=termada-data,dst=/home/termada/.config/termada \  ghcr.io/islomzoda/termada
# Homebrew:brew install Islomzoda/tap/termada
# From source (needs Go 1.26.5+):TERMADA_FROM_SOURCE=1 ./install.sh# or:  go build -o ~/.local/bin/termada ./cmd/termada

Quick start

termada serve                    # start the daemon; prints a tokenized dashboard URLtermada dashboard --open         # print and open a fresh tokenized URL

The dashboard bootstrap stores the token in browser session storage and removes it from the address bar. Static assets are public on loopback, but every TCP /api/* request and /metrics requires the token. The legacy dashboard.local_trust setting is deprecated and does not bypass API auth.

Connect it to your agent — this is a one-time, user-wide step. You do it once per user account and every project gets Termada automatically; there's nothing to copy into each repo.

For Claude Code, one command does it:

claude mcp add --scope user termada -- termada serve --stdio

Using a different agent (or prefer a file)? Add this once to your global MCP config — see .mcp.json.example:

{ "mcpServers": { "termada": { "command": "termada", "args": ["serve", "--stdio"] } } }

That's it. From now on, just ask the agent to do terminal work — it flows through Termada while you watch and control it live. The one shared daemon serves every project and every agent session, all on the same dashboard.

Configuration is strict: unknown/unsupported fields fail startup. Values written as ${NAME} are expanded from the environment, and an unset referenced variable is an error rather than an empty credential. Leave optional credential fields out until their variables are available.

Reach remote servers through Termada

For the agent to operate a remote box through Termada (observable, reconnecting, policy-gated) instead of shelling out to raw ssh, register the server once — then it's reachable by name, no IP and no raw ssh client.

Add it to config.yaml (see config.example.yaml) and restart the daemon:

servers:  - name: prod    host: prod.example.com        # hostname or IP of your server    user: deploy    # auth is OPTIONAL: a vault entry name for a Termada-stored credential.    # Omit it to try SSH_AUTH_SOCK plus unencrypted ~/.ssh/id_ed25519,    # id_ecdsa, or id_rsa. Termada does not interpret arbitrary ssh config.    # auth: prod-ssh-key    tags: [web]

…or add it live from the dashboard (Servers → Add). Confirm it's registered:

termada servers          # lists registered servers by name

Now the agent reaches it by name:

• a remote shell session — session_create(target="prod"), then run exec_run / exec_start in that session (state persists while connected; after a link drop Termada reconnects a fresh shell, but an orphaned remote process may still be running);
• one command across servers — fleet_run(command=[...], servers=["prod"]) (or by tag).
• a loopback TCP tunnel — port_forward(server="prod", remote_host="127.0.0.1", remote_port=5432), then close its returned id with port_forward_close when done.

Make the agent actually use Termada

Agents like Claude Code and Cursor ship with a built-in shell and will reach for it (and for raw ssh) by default. Two things steer them to Termada:

1. Install the usage skill — the plugin below, or skills/termada. It teaches the agent how to drive the tools (and to route remote work through registered servers instead of ssh).

2. Add a project rule so the agent prefers Termada. Put this in CLAUDE.md (Claude Code), .cursor/rules (Cursor), or your agent's system prompt:

   Use the Termada MCP tools for all shell and remote work — exec_run / exec_start for commands, session_create(target="<server>") and fleet_run for remote servers. Do not use the built-in shell or a raw ssh client: everything must go through Termada so it is observable, reconnecting, and policy-gated. If a server isn't in server_list(), ask me to register it rather than falling back to ssh. For a concrete operational outcome, start with mission_create, use its session for every command, attach real successful job ids to plan steps, and finish with mission_report.

/plugin marketplace add Islomzoda/termada/plugin install termada@termada

MCP tools

Commands are passed as an argv array (["echo", "hi"]), never a shell string, so metacharacters are literal unless you explicitly invoke a shell such as ["bash", "-lc", "..."]. Explicit shell payloads still pass through policy; ambiguous compound payloads fail closed when deny/confirm rules are present.

CLI

termada serve [--stdio]              daemon, or the MCP shimtermada dashboard [--open]           print / optionally open the dashboard URLtermada top | status                 live TUI / overviewtermada jobs [-f] | sessions         list jobs / sessionstermada logs <job> [-f]              stream a job's outputtermada kill <job> | stop            kill one job / signal all admitted worktermada pending                      list human-in-the-loop approvalstermada approve|deny [--reason text] [--ticket id] <action_id>termada audit [verify]               audit feed / verify active + rotated chaintermada servers | unlock             remote inventory / unlock the vaulttermada vault init|set|list|rm|reset manage credentialstermada snapshot create|list|restore local-FS safety net (undo)termada setup                        install user-scoped Claude MCP configtermada doctor                       health checktermada service install|uninstall|status  run the daemon at logintermada update                       Unix self-update; Windows prints manual path

Documentation

• docs/SECURITY.md — threat model: what's protected and what isn't.
• docs/ACTION_CONTROL_PLANE.md — closed action catalog and production invariants.
• docs/PLUGINS.md — writing out-of-process tool plugins.
• docs/PUBLISHING.md — release & MCP-registry process.
• docs/build-week/SUBMISSION.md — OpenAI Build Week submission package.
• docs/build-week/ARCHITECTURE.md — Mission Control architecture and trust boundaries.

Architecture

A single daemon owns runtime state; agents connect through a stdio shim, and runtime inspection and controlled actions in the dashboard, TUI, and CLI use the same daemon control plane. Host-owner maintenance commands such as vault-file creation/reset, installation, service registration, self-update, audit-file verification, and configuration editing run outside that API and remain part of the trusted OS/deployment boundary.

cmd/termada            CLI: daemon, shim, inspection/control, vaultinternal/engine        sessions, jobs, PTY, status machine, signals, files, recipesinternal/action        immutable Principal + Action + Resource + Context contractinternal/control       authorization, approval, budget, intent and outcome gatewayinternal/ledger        durable one-use grants, budgets and export outboxinternal/budget        typed policy-owned budget resolutioninternal/adapter       typed prepare/execute/verify/compensate/reconcile SDKinternal/identity      pinned-JWKS human and machine JWT verificationinternal/anchor        signed external audit receipts and durable deliveryinternal/output        cursor buffers, VT cleaner, redactioninternal/policy        argv allow/deny/confirm classificationinternal/vault         age-encrypted credential storeinternal/audit         hash-chained tamper-evident loginternal/bus           event bus (observability + durable audit)internal/daemon        long-lived process: listeners, auth, lifecycleinternal/controlplane  HTTP/JSON control-plane server + clientinternal/dashboard     embedded web UIinternal/tui           termada topinternal/fleet         server selection + concurrent aggregationinternal/sshx          SSH runner (vault / agent / key auth, TOFU host keys)internal/mcp           MCP JSON-RPC stdio server + toolsinternal/mission       durable missions, evidence validation, recovery, reports

Development

make vet test    # vet + testsmake race        # tests under the race detector./demo/mission-control/verify_demo.sh  # real MCP + approval + HTTP 503→200 flow

Engine tests exercise a real PTY and bash; fleet logic is unit-tested with a mock runner; the daemon stack and SSH are integration-tested end-to-end.

License

Apache-2.0.