# BaseCode – MDM for Coding Agents > Source: > Published: 2026-07-28 09:05:28+00:00 # Know and control how coding agents are being used in your organization. Your engineers run **Claude Code, Codex, Copilot, Gemini and opencode** on their machines. basecode lets you centrally inject **agent identities, coding policies, skills, MCP servers and models** into all of them — then **observe every prompt, project and token**, and **enforce guardrails and budgets** across the fleet. | LangGraph concept | Lyzr ADK equivalent | | Graph nodes/edges for multi-step flow | Multiple studio.create_agent() instances, orchestrated via sequential .run() calls | | Shared state across nodes | Context objects or session memory | | Tool-calling nodes | agent.add_tool() with typed Python functions | | Persistence/checkpointing | memory=N + session_id on agent.run() | ### Employees leave. Knowledge stays. Every project's context, decisions and learnings are captured in a **knowledge graph on your gateway** — not in one person's head. Attrition doesn't take your codebase's memory with it. ### Nothing is a black box. Every prompt, token, model call and tool call across the fleet is **observed, attributed and audit-logged** — so what your agents build is never a mystery. ### Day-one onboarding. A new hire's agent inherits **profiles, skills, MCP servers and project knowledge** the moment their device enrolls — they build like a veteran from the first prompt. **harness-agent-service** 53h 21m · 97.3M tok **flabot** 23h 43m · 63.0M tok **gitview** 6h 35m · 23.1M **vite-architect-core** 5h 06m · 10.2M tok **hyoconpute** 2h 47m **glupgorton** 14h 03m · 36.3M tok **architect-studio** 4h 47m **desktop** 1h 13m **agent-studio-ui** 2h 11m **proxy** 58m **docs** 41m **3** Owners **34** Devices **7d 29h** Coding time **601.2M** Tokens ## Who is building what — across every person, project and device. basecode attributes real coding activity — **active coding time, recency and token spend** — across **Claude Code, Codex and Copilot**, rolled up per person or per group. Duplicate device snapshots of the same folder are counted once, across all of a person's devices. - ✓ **Per-person or per-group treemaps**— pivot by coding time, 7-day activity, or token spend. - ✓ **Agent attribution**— every project shows which agents built it: cc, cdx, cplt badges. - ✓ **Zoom out to the whole org**— regions group the same project across owners; click a tile to open that person. **1,284** Requests **2,845ms** Avg latency **134.8k** Tokens **US$0.14** Cost ## See every prompt. Guard every response. Every request from every employee's coding agent flows through the gateway — and shows up in a **live log** with latency, tokens and cost. Enforcement runs **in the request path, at the gateway** — it can't be switched off from a laptop — and every decision is retained as an audit-ready record. - ✓ **Policy-engine guardrails**— evaluate prompts and responses against org policy with OPA / Cedar: allow, deny or redact before anything reaches a model. - ✓ **Slice by anything**— status, model, provider, virtual key, routing rule, user, team or customer. - ✓ **MCP logs too**— every connector interaction observed and audit-logged. ## Remotely swap the model inside any coding agent. Cut cost with complexity routing. Route **Claude Code, Codex and Copilot through the basecode gateway** — or let any of them keep their own subscription. Map their model slots to **any provider's model**, hosted or self-hosted, behind scoped **virtual keys**. - ✓ **Complexity Router**— classifies each request and routes simple edits to cheap models, hard reasoning to frontier ones. Routine work runs at a fraction of frontier-model cost. - ✓ **Hard budgets**— spend ceilings per group, per model, per employee,** enforced before the request is sent**, not an alert after the invoice arrives. - ✓ **Virtual keys & scopes**— agents never hold raw provider keys. Off-scope models get 403 Model not allowed. ## Give every coding agent an identity your company controls. An Agent Profile is a **git repo describing an agent** — its rules, soul, duties, knowledge and skills. Profiles live on the gateway per group and **auto-sync to every member device**. Whatever agent your developers run, it follows the profile. - ✓ **Just git — versionable governance.** Update enterprise knowledge, coding policy and rules with commits, branches, PRs and rollbacks. Every policy change is diff-able and auditable. - ✓ **A defined, open structure.** agent.yaml · RULES.md · SOUL.md · README.md + knowledge/ and skills/ folders. No proprietary lock-in. - ✓ **Auto-synced to the fleet.** Stored on the gateway per group, pushed to member devices on join and on every change. - ✓ **Policy the agent actually enforces.** Framework restrictions, coding standards, observability instructions — the agent refuses off-policy work and redirects to your approved stack. **Approved-stack policy, checkbox by checkbox** — admins pick exactly which coding, frontend, backend and agent frameworks, deployment targets and observability tools every agent may use. Agents inherit the selection and **refuse anything off-list**. **lyzr-adk-builder:** builds exclusively with the Lyzr ADK — LangChain, CrewAI, AutoGen and raw API calls are denied. One repo defines it; the whole fleet obeys it. ## One hideable TUI that drives every coding agent — in any IDE. **Your developers keep the agents they already love, exactly as they are** — no rip-and-replace, no adoption cliff. basecode ships a hideable client TUI (F2 to toggle, Ctrl-] to detach and keep running) that wraps the agents you already use — **100% additive**, compatible with any IDE or terminal. ## Swap a session between agents in one click. Carry the full history and context of a conversation from Claude Code into Codex, Copilot or any other agent — **mid-task, no copy-paste**. - ✓ **Skills & MCP injection**— author skills centrally, approve MCP servers once, and one-click install them into anyone's coding agent from a synced catalog. - ✓ **Ship custom subagents to every developer**— plain-markdown definitions written to .claude/agents/ on every device. One meticulous code reviewer, org-wide, in one save. - ✓ **Broadcast to everyone, assign work to anyone**— fleet-wide banners land inside every agent; agent-enforced TODOs report ticks attributed to enrolled email. - ✓ **No coding agent yet?** basecode ships its own —**Architect**, a built-in Rust coding agent: very low memory, fast edits, expert at building agents. ## Cost control, fleet ops and a built-in agent — without the rip-and-replace. ### Complexity Router Classifies each request and routes simple edits to cheap models, hard reasoning to frontier ones — automatically. Routine work runs at a fraction of frontier-model cost. ### Groups Device cohorts with scoped everything — skills, MCP, profiles and subagents auto-sync to every device that joins. Manager vs admin authority, presence built-in. ### Broadcasts & tasks A live channel to your whole engineering fleet — announcements land as banners in every TUI, and agent-enforced TODOs report completion per enrolled email. ### Architect, built-in basecode's first-party coding agent, written in Rust — very low memory, fast edits, and exceptionally good at building agents. ## Built on an open standard — the Git Agent Protocol. Every basecode profile is an **OpenGAP repo** — the open format for describing agents as versionable git, maintained by **Lyzr** and adopted by a growing open-source community. Your agent governance is portable, diff-able and vendor-neutral by construction — **if you ever leave basecode, you take all of it with you**: profiles, policies and knowledge are plain git. ## Everything basecode does — in one view. You've seen each capability up close. Here are all twelve together, ordered the way a buying committee reads: **see** everything, **control the spend**, **govern** the work, and roll out with **zero disruption**. ### Fleet observability **Coding time, recency & token spend** per person, project and device- Treemaps per person or per group, with agent-attribution badges - Duplicate folder snapshots counted once across devices ### Guardrails & LLM logs - Every prompt logged with **tokens, latency and cost** - OPA / Cedar policies allow, deny or redact in-line - MCP tool calls get the same audit treatment ### Budgets & limits **Hard spend ceilings** per group, per model, per employee- Token & request rate limits with hourly / daily / monthly resets - Enforced at the gateway — not on the honor system ### Model gateway - Remap any agent's model slots to **any provider**— hosted or self-hosted - Scoped **virtual keys**— agents never hold raw provider keys **Complexity Router:** cheap models for easy tasks, frontier for hard ones ### Agent Profiles (OpenGAP) - Each agent's identity is a git repo — agent.yaml, RULES.md, SOUL.md, knowledge & skills - Auto-syncs to every member device on join and on every change - Agents **refuse off-policy work**— no plugin, no fork ### Approved-stack policy - Six dimensions: coding, frontend, backend, deployment, observability, agent framework **Checkbox governance** per workspace or per group- Agents inherit the selection and refuse anything off-list ### Client TUI - Hideable sidebar (F2), detach and keep running (Ctrl-]) **100% additive**— wraps the agents you already use, in any IDE- Swap a session between agents mid-task, full context carried ### Skills & MCP injection - Author skills centrally — scoped by agent compatibility and allowed tools **Approved MCP catalog:** one-click install into anyone's agent- Pushed per group or per employee, synced live ### Subagents - Plain-markdown agent definitions — **no SDK, no build step** - Written to .claude/agents/ on every device - One code reviewer pushed org-wide in one save ### Groups - Device cohorts — skills, MCP, profiles and subagents **scoped per group** - Manager vs admin authority; safe removal semantics - Presence built-in: device counts, online status, last seen ### Broadcasts & tasks - Fleet-wide banners in every TUI — live delivery + offline history **Agent-enforced TODOs** with ticks attributed to enrolled email- Everything audit-logged ### Architect + OpenGAP **Built-in Rust coding agent**— very low memory, fast edits- Exceptionally good at building agents (Lyzr ADK workflow) - Profiles are portable **OpenGAP repos**— an open standard ## Your engineers already adopted coding agents. Now govern them. Identity, skills, MCP, models, budgets, guardrails, observability and fleet ops — **one control plane**, synced live to every device. [shreyas@lyzr.ai](mailto:shreyas@lyzr.ai)