Skip to main content
AgentIDEAgentIDE in the source-owned AgentIDE documentation.AgentIDEreferenceagentidereferenceevaluatoradopterdeveloperoperatorbuild-agent-systemsevaluate-beyond10x-productsreference

AgentIDE

AgentIDE is a local coding workbench where humans and agents share one durable session. Files, diffs, terminal activity, model actions, evidence, and approval decisions remain visible and replayable instead of disappearing inside a chat transcript.

AgentIDE local browser workbench showing a new session, open-file and approval regions, and the durable session timeline

AgentIDE is a developer preview for Linux x86_64. The CLI, model-backed terminal UI, projection-only terminal UI, and loopback browser all use the same session record.

Install AgentIDE​

Install the latest stable GitHub Release with the first-party installer:

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/beyond10x/agentide/releases/latest/download/agentide-installer.sh | sh

The installer resolves an immutable release, downloads its Linux x86_64 archive and SHA256SUMS, verifies exactly one matching checksum, and installs agentide into ~/.local/bin without sudo. If that directory is not on PATH, it prints the required next step.

agentide --version

Run your first coding session​

Run these commands from the workspace you want AgentIDE to use. The endpoint, exact model, and credential source are intentionally explicit; AgentIDE does not choose a provider or search for credentials.

export AGENTIDE_BASE_URL="https://api.example/v1"
export AGENTIDE_MODEL="model-id"
export MODEL_API_KEY="your-api-key"

agentide run \
  --objective "Implement and verify the change" \
  --api-key-env MODEL_API_KEY

agentide run adopts the current directory, creates a durable session, prints its ID and resume command, and opens the model-backed TUI. The session remains available after the TUI exits or the model loop fails. The key is read from MODEL_API_KEY for each request and is never written to the session journal.

Inside the TUI, press i to prompt the agent, Ctrl+K for the command palette, Ctrl+P to open a file, and Tab to move between regions. When an effect needs approval, y or n resolves the exact plan displayed by the gate.

Choose a running mode​

ModeStartBest for
Model-backed TUIagentide run --api-key-env MODEL_API_KEYInteractive agent-assisted coding
Projection-only TUIagentide runInspecting and operating a session without a model
JSON CLIagentide session start, then agentide intent ...Scripts and agent tool adapters
Local browseragentide serve --session <id>A loopback visual view of the same session
Hosted DevCenterEnvironment-provided URLAuthenticated composed deployments; no public self-service endpoint

For an existing session, attach directly with agentide tui --session <id>. Model-backed run and tui require both AGENTIDE_BASE_URL and AGENTIDE_MODEL; supplying only one is an error. Omitting both selects projection-only mode. See the generated running-mode reference for the exact implementation, attachment, availability, and ESS surface of every realization.

The JSON workflow remains available when a caller needs structured output:

agentide session start --workspace . --objective "Inspect the change"
agentide snapshot --session <id>
agentide intent call --session <id> code_read --input '{"path":"src/lib.rs"}'

The browser is loopback-only. Start agentide serve --session <id>, then open http://127.0.0.1:7788/. It presents the durable projection but does not run a model loop.

The root uses the compatibility Vanilla DOM renderer. The same live session can be compared through /renderers/vanilla/ and /renderers/vue/. Both targets accept the released renderer frame and event contracts and emit typed semantic actions; HTTP, polling, routing, authentication, storage, and sockets remain in the browser host. The Vue target also exports an AgentIdeVueWorkbench composition shell with named regions for host-owned editors, diffs, terminals, and inspectors. It owns layout and landmarks without acquiring access to a host API.

Install the latest source with Cargo​

If you deliberately want the newest source from main and have a Rust toolchain, install directly from GitHub:

cargo install --git https://github.com/beyond10x/agentide --locked agentide-cli

This command compiles the current main, which may be newer than the latest stable GitHub Release. It does not select a release tag. AgentIDE is not published to crates.io, so cargo install agentide-cli is not a supported command. Use the installer above when you want the latest stable prebuilt binary.

Why AgentIDE is built on ESS​

ESS turns AgentIDE's coding model into validated intermediate representation and generated contracts. That matters to a user because:

  • an action such as code_read, code_edit, or code_verify keeps the same meaning across CLI, TUI, browser, and hosted surfaces;
  • the model sees typed semantic actions rather than being allowed to invent executables, credentials, destinations, or policy;
  • effectful work is previewed as an exact digest-bound plan before approval;
  • sessions and evidence can be replayed against the same declared behavior;
  • the generated running-mode reference cannot silently disagree with the executable specification.

Harness drives the model loop and Substrate supplies guarded workspace implementations. Service SDK and Eventlog provide the generated hosted form without introducing a second AgentIDE-specific service or persistence model.

Safety model​

Session state lives under ${XDG_STATE_HOME:-$HOME/.local/state}/agentide, outside the target workspace. A model request cannot select a driver, executable, credential, destination, or policy. Missing Substrate facts, bindings, profiles, approvals, or recovery facts produce named refusals; AgentIDE does not fall back to unconfined host effects.

Mutating JSON intents use two phases: preview the exact plan, grant its SHA-256 digest, and resume with the same input. The TUI presents the same approval boundary interactively. Model conversations and credential values are not written to the AgentIDE journal.

Learn more​

Contributors can run the complete repository gate with cargo xtask gate. After changing ESS or Service SDK definitions, regenerate only through the repository-owned xtask commands described in AGENTS.md.