A lean, container-isolated, memory-aware multi-agent runtime
OpenLegion provides 9 tools with explicit schemas, descriptions, and parameter types. Most tools follow verb_noun naming (browser_navigate, browser_warmup, hand_off, team_drive, run_command, execute_code, read_file, write_file, list_files). Descriptions are detailed and contextual (avg ~250 chars), exceeding the 194-char baseline. All parameters have type definitions and descriptions. However, several tools lack output schema documentation, error handling guidance is minimal, and some parameter constraints are informal (e.g., 'max 6,000 chars' in brief param is stated in description but not enforced via schema). Tool composition is strong, tools chain well (e.g., list_files → read_file). Security is well-handled: no credentials in params, destructive operations documented (hand_off, team_drive, run_command, execute_code marked WRITE/IRREVERSIBLE). Risk annotations present but not formalized as toolAnnotations in the MCP sense.
Navigate your browser to a URL and return the page text. Use this to visit any website: sign-up pages, dashboards, search engines, web apps. You can then use browser_get_elements to get element refs, then browser_click and browser_type with refs to interact. For heavy SPAs (X/Twitter, Gmail, etc.) use wait_until='networkidle' or wait_until='load' so the page fully renders before you read elements. Set snapshot_after=true to include element refs in the response (saves a separate browser_get_elements call). Leave 'referer' unset for normal browsing — the service picks a plausible referer (search engine, same-origin, etc.) based on the destination so the page sees a realistic arrival pattern. Override only when you're explicitly modeling a specific click-through.
Pre-task warmup: navigate to a search engine and the target's home page before starting the main task. Use this once at the start of an automation session against LinkedIn / X / Twitter / Facebook / Instagram (and other behavior-fingerprinting targets) so the session has a realistic browsing trail. NOT needed for casual sites without bot detection. Do NOT call this multiple times per session — once at the start is enough; the session already has a referer history after that. Returns a structured summary of the steps that ran so the agent can log what happened.
Run a Python snippet in your Linux environment and get back ONLY what you print(). Use this to collapse several steps into one turn: fetch, compute, filter, and print just the final answer — intermediate values never come back to you, only your printed output. The snippet runs as a fresh `python3` process (no state carries over between calls). On failure you get the error (stderr) and a non-zero exit code. print() what you need; everything else is discarded.
Output schemas not documented. Tools like browser_navigate, hand_off, team_drive, run_command, execute_code lack explicit return type definitions. LLMs cannot plan downstream tool calls or extract structured data without knowing what fields to expect.
team_drive action parameter uses enum but lacks clear guidance on when to use each action. 'submit_review' and 'record_verdict' have complex preconditions (team membership, lead role) not documented in the parameter description.
Error handling is minimal. Tools like run_command and execute_code state they return 'error (stderr) and non-zero exit code' but do not guide the LLM on recovery (retry, ask user, abort). No error classification (retryable vs fatal).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 70 | 2026-07-28+ | v2 |
Hand off work to a teammate. Creates a task in their inbox and wakes them up automatically. This is your PRIMARY coordination tool — use it whenever you've completed work that another agent should act on. Keep the 'summary' SHORT — it becomes the task title in the recipient's inbox and on the dashboard. Aim for ≤80 characters, like a Git commit subject line. Examples: 'Draft Q3 launch brief' or 'Review pricing change for SKU-123'. For any non-trivial handoff, ALWAYS pass 'brief'. The recipient does NOT see your conversation — they start from only what you send, so a one-line summary with no brief produces shallow work. Write the brief as markdown covering: the objective, background the recipient lacks (who asked and why, constraints, what you already know), success criteria, and the expected deliverable (form, depth, where to save it). The brief is delivered to the recipient together with the task. If you have structured output data to share, pass it as 'data' (JSON string). It is committed to your Team Drive and the task carries a drive:// pointer the recipient reads via team_drive. (If you are not on a team, the data is folded inline into the brief instead.) For lightweight handoffs where the summary is enough context, omit 'data' and 'brief'. The target agent sees the task in their inbox (via check_inbox) with your summary, your brief, and a pointer to your output. hand_off delegates WORK. If you only need information from a teammate, use ask_teammate instead — it returns their answer inline without creating a task. If a recipient blocks a task you created with a question, you'll see it in check_inbox() events[]: answer it with ask_teammate(to=<recipient>, ...) so they can resume the SAME task — do not re-issue a duplicate hand_off.
List files and directories in /data. Returns names, sizes, and types (file/dir). Use pattern for glob filtering (e.g. '*.py') and recursive=true to search subdirectories.
Read a file's text content from the /data directory. Returns the content as a string. Use offset and limit to read large files in chunks. Returns an error if the file does not exist — use list_files to check first.
Run a shell command in your Linux environment and return stdout+stderr. You have Python, Node.js, curl, git, and standard tools. Use for: installing packages, running scripts, processing data, downloading files, or any system operation. Do NOT use run_command to modify workspace identity files (SOUL.md, INSTRUCTIONS.md, HEARTBEAT.md, USER.md, MEMORY.md) — use update_workspace instead.
Your team's shared git repository — the durable home for files you collaborate on with teammates (documents, code, data). It lives on the mesh; your working copy is /data/drive. Workflow: team_drive('clone') once, then team_drive('branch', branch='my-feature') to start work, edit files under /data/drive with your normal file tools, team_drive('sync', message='what changed') to commit+push, and team_drive('submit_review', title='...', summary='...') when the branch is ready to integrate. The 'main' branch is integrate-only — you CANNOT push to it; a human merges approved reviews. 'pull' refreshes from the drive, 'list_reviews'/'log'/'status' inspect state. 'record_verdict' records YOUR advisory approve/reject opinion on an open review (review_id, verdict, optional note) — LEAD-ONLY, enforced server-side (403 if you aren't your team's lead). It has no effect on merge/reject; a human still decides those. Solo agents (not on a team) have no Team Drive.
Write content to a file in the /data directory. Overwrites the file if it exists; creates parent directories if needed. Use append=true to add to an existing file. For deliverables you want to share with teammates or the user, use save_artifact instead.
Parameter constraints stated informally in descriptions rather than enforced via schema. E.g., hand_off brief 'max 6,000 chars', summary '≤80 characters' are prose constraints, not JSON Schema maxLength. LLMs may ignore prose limits.
browser_navigate referer parameter description is verbose and tutorial-like rather than concise. At ~300 chars, it exceeds best practice (50-150 chars per param). Simplify to: 'Optional referer URL. Omit for realistic auto-detection; pass empty string for no referer.'