● Open Source India 2026
The EnterpriseContext Problem
Legacy APIs to AI agents: building the MCP layer for enterprise software
Sandesh Bandal
Developer Advocate
Niranjan Akella
Engineering Manager, AI/ML
Built by
SANDESH · 0:00 – 0:30
"Every company has been told to adopt AI agents. Today we want to talk about the thing that quietly decides whether those agents are useful: context."
Sandesh opens: introduce Sandesh (Developer Advocate) and Niranjan (Engineering Manager, AI/ML) from 2nd Brain Labs, the team behind Synapse. Promise: the problem (6 min), how to build the MCP layer (7 min), a live demo (7 min), then questions.
01
Search stopped returning links.answers .
Go June to August, when the waterfalls are at their wildest.
Ride in to Madikeri, trek Tadiandamol, stay on a coffee estate. Pack leech socks.
trek blogs reddit thread maps · reviews +11
2.5B+
people a month read Google's AI Overviews. A question fans out into many searches; an AI system reads them and answers, with sources.
Google I/O, May 2026
SANDESH · ~1 min · 2 clicks
"Think about the last time you searched for something. You didn't get ten blue links. You got an answer."
Click: the answer assembles. Information from every direction is retrieved, reasoned over, and cited. Click: 2.5B+ people a month already use this. Knowledge plus reasoning is how we cut hallucinations and get to the right point. "Context" is now everyday vocabulary. Everyone expects an answer, not a search. Transition before the click to the next slide: "Now let's think about this in the context of your company." AI Overviews 2.5B+ MAU, AI Mode 1B+ MAU: Google I/O 2026 keynote (blog.google, May 19 2026). "Query fan-out": Google, May 2025.
01
Gone are the days of manual search.
Agents speak our language. What they say is only as good as the knowledge they can reach.
History
Search
Keywords in, ten links out. You read, collate and decide.
Yesterday
Retrieval + LLM
It reads for you and answers in plain language, with sources.
Today
Agents
It reads, reasons and acts across your systems.
The bottleneck moved from search to context .
01
Every major lab is going fully agentic .
OpenAI Sep 2026 dots
Always-on ChatGPT agents, each with its own computer.
Anthropic 2025–26 Claude Code & Cowork
Agents that code, browse and work across your apps.
Meta Sep 2026 Muse
A personal AI agent, built on Muse Spark.
xAI Aug 2026 Grok Bot
Persistent agent teammates on their own cloud computers.
Open source MIT · 391k+ ★ OpenClaw
A personal agent you run yourself, across 20+ chat apps.
Open source MIT · 251k ★ Hermes Agent
Nous Research's agent that writes its own skills.
They all hit the same wall: your company's context .
SANDESH · ~1 min · 1 click
Land first, before reading anything: "Now let's put this inside your company. What is happening on the enterprise side?" Then: this isn't one vendor's experiment any more. In the last few weeks alone: OpenAI's dots (DevDay, 29 Sep): always-on ChatGPT agents with their own computer and browser. Anthropic: Claude Code, Cowork, Managed Agents, Claude in Chrome. Meta's Muse (8 Sep), a personal agent on Muse Spark. xAI's Grok Bot (11 Aug): persistent agent teammates on cloud computers. And the open-source side is just as loud: OpenClaw (MIT, 391k+ stars) and Nous Research's Hermes Agent (MIT, 251k stars). Click: brilliant agents, all with the same blind spot: they can't see inside your company. OpenAI DevDay / TechCrunch 29 Sep 2026 (dots) · claude.com (Cowork Jan/Apr 2026, Managed Agents Apr 2026, Claude in Chrome GA Aug 2026) · about.fb.com 8 Sep 2026 (Muse) · x.ai 11 Aug 2026 (Grok Bot) · github.com/openclaw/openclaw, github.com/NousResearch/hermes-agent (stars as of 4 Oct 2026).
01
Inside every enterprise, the knowledge is locked in systems agents can't reach.
SANDESH · ~1 min · 2 clicks
"Now take that agent inside a company."
The knowledge that matters lives in ERPs, warehouse systems, EHRs, CRMs and 25-year-old AS/400 boxes. They have endpoints, but nothing an agent can speak to. Click: the agent reaches out and hits a wall at every system. Optional, not on screen: the average enterprise runs 957 apps and only 27% are integrated (MuleSoft 2026). Legacy isn't going away: 95% of organisations are keeping or increasing mainframe use (Kyndryl 2025). MuleSoft / Salesforce Connectivity Benchmark 2026 (Feb 5 2026; 1,050 IT leaders). Kyndryl State of Mainframe Modernization, Sep 2025.
01
Without context, a model guesses. With context, it answers .
Without context
Which plant has brake caliper brackets in stock right now?
Detroit usually keeps a buffer of these parts. I'd check there first.
Guess · no source
With context
Which plant has brake caliper brackets in stock right now?
find_part_across_facilities (part="BRK-4471")
Stuttgart has 280 available (340 on hand, 60 committed). Puebla needs 120.
✓ Source · AS/400, Stuttgart
SANDESH · ~45 s · 1 click
Left: same model, no access. It sounds confident and it's wrong, or it refuses. That's a hallucination with good manners. Click: give it one tool on one real system. Now it answers with a number and a source. Grounding isn't a nice-to-have. Retrieval measurably reduces hallucination, but only if the agent can reach the source. Shuster et al., "Retrieval Augmentation Reduces Hallucination in Conversation", EMNLP Findings 2021. Stanford HAI 2024: general chatbots hallucinated 58–82% on legal queries.
01
Every agent × every system is a custom integration .
6 × 8 = 48 connectors
Each one: read the code, write middleware, secure it, host it. Weeks each.
01
The enterprise context problem is reach , knowledge and trust .
01
Reach
Agents can't get into the systems where the business actually runs.
02
Knowledge
What a system does, and how it relates to the rest, lives in people's heads and scattered docs.
03
Trust
Who may see what? An agent must inherit a person's access, never exceed it.
Solve all three, and an agent stops guessing and starts answering .
SANDESH · ~45 s · HAND OFF at the end · 4 clicks
Reach: the plumbing. Knowledge: the understanding. Trust: the governance. Gartner expects 40% of enterprise apps to have task-specific agents by 2026, and over 40% of agentic AI projects to be cancelled by 2027. Missing context and weak risk controls are why. "So how do you actually build this layer? Over to Niranjan." Gartner press releases: Aug 26 2025 (40% of apps); Jun 25 2025 (40% cancelled by 2027).
Introducing
The Enterprise Context Engine
NIRANJAN · 6:00 · the big moment, 15 s
"If you solve reach, knowledge and trust, you solve the enterprise context problem. That's exactly what we're building. Introducing Synapse, the Enterprise Context Engine."
Let the logo land. Pause before talking.
02
Synapse solves reach , knowledge and trust .
MCP servers today. Every context surface your AI stack needs, tomorrow.
Reach Connect
MCP servers from code, Swagger and APIs, hosted for every agent
Today · open source Knowledge Knowledge fabric
systems, docs, drives and recordings in one knowledge layer
Next Trust Governance
who can see what, enforced at the source · multi-tenant
Next
02
Synapse is the bridge between every enterprise system and the AI world.
Legacy
AS/400 · RPG programs SAP ECC · BAPI / RFC Mainframe · CICS, MQ SOAP services
Modern
REST · OpenAPI GraphQL SaaS APIs Internal microservices
Context layer
auth secrets audit
The system stays untouched. The AI handles the rest.
02
Synapse CLI is open source . Any codebase becomes an MCP server.
Weeks → Days
From an integration project to live tools.
Runs client-side
The agent asks for the bytes it needs. It never sees your whole codebase.
Bring your own model
--local with Anthropic, OpenAI, Groq, xAI, OpenRouter, Ollama or your endpoint.
Auto mode
Detects FastAPI, Flask, Django, Express, NestJS, Spring, Gin, ASP.NET, Axum.
Auto mode today Python TypeScript JavaScript Go Java .NET Rust C, C++ and more coming
NIRANJAN · ~1 min · plays on its own
This is the open-source part, and the reason we're here: Apache-2.0, on GitHub and npm. Integration that took weeks now takes days. Install from npm (synapse init once for the API key). analyze : an Aho-Corasick prefilter sniffs the first 64 KB of every file so only files with routes or functions reach tree-sitter; a parallel parse pool extracts routes, functions, classes and models into a deterministic surface manifest (same repo → same manifest, no LLM involved). It scales to 500k-line repos.build --auto : detects the framework and turns every route into a tool, registered live on the Synapse runner, so it is immediately usable from Claude, Cursor or ChatGPT.How build works: a streaming session where the agent asks the CLI for exactly what it needs (read_file, grep, find_definition). Exploration stays on your machine. Auto mode for REST frameworks; Custom mode composes tools from any functions you describe in plain English. Python and TypeScript are first-class; Go, Java, C#, Rust have route detection. Terminal output is illustrative (counts from the Synapse site demo). CLI v0.1.5, Apache-2.0: github.com/2ndbrainlabs-ai/synapse-cli-ts
02
Every endpoint in your Swagger is a tool waiting.
Inventory API 1.0.0 OAS 3.1 /docs
GET /parts/{part_id} Get a part ✓
GET /stock Stock at a facility ✓
GET /parts/search Search parts ✓
POST /transfers Create a transfer ✓
PATCH /transfers/{id} Update a transfer ✓
DELETE /reservations/{id} Cancel a reservation ✓
MCP server
Hosted, every host can call it.
or
Agent skill
The same tools, packaged with know-how.
Reads today codebases Swagger / OpenAPI cURL endpoints by hand Next docs · wikis · Postman · GraphQL · gRPC
02
Synapse hosts your MCP server, so every agent can reach it.
Console · Manage MCP Context
GET /stock/{part_id}
GET /parts/search
POST /transfers
Import from Swagger Generate MCP Server
↑ Deploy to Synapse Cloud
Self-hosted · your cloudnext
// claude_desktop_config.json
"mcpServers": {
"inventory": {
"command": "npx",
"args": ["-y",
"@2ndbrainlabs-ai/synapse-mcp "],
"env": {
"SYNAPSE_SERVER_ID": "…"
}
}
}Secrets resolve at runtime and never reach the model. Every tool call is metered and audited.
02
Synapse CLI is open source. Help it read everything .
github.com/2ndbrainlabs-ai/
Apache-2.0 · fork it · explore it
Wanted most
Read from docs
API docs, Markdown and wikis become tools.
Formats
More specs
Postman collections, GraphQL schemas, gRPC protos.
Frameworks
New route detectors
Teach auto mode Laravel, Rails, Phoenix and more.
Languages
Deeper analysis
Full Java, .NET, Go and Rust support.
$ npm install -g @2ndbrainlabs-ai/synapse-cli
03
“Line 3 in Puebla is stopping. Where else do we have the part ?”
Tool calls
find_part_across_facilities ("BRK-4471")
Stuttgart · 280 available
create_transfer_order (120, Stuttgart → Puebla)
✓ TR-0001 reserved
create_transfer_order (500, Chennai → Puebla)
Refused: only 40 available
NIRANJAN · 13:00 – 20:00 · LIVE DEMO · 3 clicks = backup if the live demo fails
"Now let's put it into action."
Run of show (switch to the demo apps):
1 · CLI (2 min): in a FastAPI repo, run synapse analyze then synapse build ; show the generated tools.2 · Console (2 min): Import from Swagger → Generate → Deploy to Synapse Cloud; copy the config into Claude.3 · Agent (3 min): in Claude, ask "Line 3 in Puebla is stopping: we're out of brake caliper brackets. Where else do we have them?" → "Move 120 from Stuttgart to Puebla" → guardrail: "Pull 500 from Chennai" is refused (only 40 available).Five plants, five systems of record: Puebla (SAP ECC 6.0), Detroit (Oracle WMS), Stuttgart (AS/400), Chennai (custom app), Shenzhen (supplier portal).
If Wi-Fi fails: stay here and click through the three steps. They tell the same story.
04
MCP is what an agent can do. A skill is how your company does it.
skills/vendor-onboarding/SKILL.md
--- name: vendor-onboardingdescription: Onboard a new vendor end to end: compliance, approval, ERP record. --- 1. Run the compliance check → kyc_check 2. Finance approves over ₹10 L → request_approval 3. Create the vendor in SAP → create_vendor
Skill · your know-how
open standard · agentskills.io Steps, policies and scripts, packaged as a folder.
MCP tools · what an agent can do
kyc_check request_approval create_vendor
Systems
KYC service · approvals engine · SAP ECC
04
Next: one knowledge layer , fed from every direction.
The right system, the right knowledge, the right access, for every agent.
Systems · via MCP Docs & wikis Drives · OneDrive, Google Drive Tickets & chats Recordings & transcripts Data & warehouses
NIRANJAN · ~45 s · 1 click · ROADMAP, say "next", not "today"
That was Reach, available today. Knowledge is next. The knowledge layer does for your company what Google does for the web: systems via MCP, plus docs and wikis, drives, recordings and transcripts, tickets and chats, data. Click: the wave becomes a globe. Every source streams into one knowledge layer that every agent can draw on, with the right access for whoever is asking.
04
Govern context at the source : agents inherit access, never exceed it.
Same question, two people
What's the Q3 compensation budget for the platform team?
Policy at the source
01 identity
02 role & policy
03 tool scope
04 data filter
05 audit log
HR director · allowed
₹4.2 Cr, 61% committed. Source: HRMS.
✓ policy hr.comp.read
Engineer · not allowed
You don't have access to compensation data.
80%
of companies say their AI agents have already taken unintended actions.
SailPoint, May 2025
Today per-user scoping · signed API keys · every tool call audited
Next org-level roles and policy-aware context
NIRANJAN · ~1 min · 3 clicks
Context without governance is a liability. The question isn't only "can the agent reach it?" but "should this person see it?" Click: the HR director asks, policy allows, answer with a source. Click: the engineer asks the same thing. Denied at the source, and logged. Click: 80% of companies say agents have already taken unintended actions. OWASP calls it "excessive agency": too much functionality, permission and autonomy. Be honest about status: today Synapse has per-user scoping, signed keys and per-call audit. Org-level roles and policy-aware context are next. SailPoint AI agent adoption report, May 28 2025 (353 IT pros). OWASP Top 10 for LLM Apps 2025, LLM06 Excessive Agency. OWASP Top 10 for Agentic Apps, Dec 2025. (The ₹ figure is illustrative.)
05
Join Synapse Stars , our new developer program.
Build projects, integrations, MCP servers
Create tutorials, videos, demos
Connect meetups, workshops, hackathons
Enable help and mentor other developers
Shape feedback, use cases, ideas
Four levels Pulse → Spark → Nexus → Nova
Badges, community, stage time, early access, rewards and API credits, internships.
synapse.2ndbrainlabs.ai/
Build. Create. Connect. Lead.
05
Agents are only as good as the context they can reach.
01 Reach, today . Open-source CLI and Synapse Cloud make any system reachable by any agent.
02 Knowledge, next . One knowledge layer, fed from every direction.
03 Trust, next . Governance at the source: agents inherit access, never exceed it.
$ npm install -g @2ndbrainlabs-ai/synapse-clisynapse.2ndbrainlabs.ai
Live demo · Booth SA-4
Your Enterprise Context Engine
Thank you · Questions? · Booth SA-4
Contribute
github.com/2ndbrainlabs-ai
Built by Try it
synapse.2ndbrainlabs.ai
02
MCP is the open standard that lets any agent use any tool .
AI host · MCP client
Claude, ChatGPT, Cursor, Gemini…
MCP server
Tools actions
Resources data
Prompts workflows
Your system
SAP · AS/400 · any REST API
Untouched. It keeps doing its job.
→ tools/call check_stock({ part: "BRK-4471", facility: "Stuttgart" }) ← 280 available
Nov 2024 · Anthropic Mar 2025 · OpenAI Apr 2025 · Google May 2025 · Microsoft Dec 2025 · Linux Foundation 110M+ SDK downloads / mo
NIRANJAN · ~45 s · 2 clicks
Model Context Protocol: a host (Claude, ChatGPT, Cursor) talks JSON-RPC to an MCP server, which exposes tools, resources and prompts. Click: a real call. The agent discovers check_stock , calls it, gets a number. Click: it's no longer one vendor's idea. Anthropic launched it in Nov 2024; OpenAI, Google and Microsoft adopted it; in Dec 2025 it moved to the Linux Foundation's Agentic AI Foundation. 110M+ SDK downloads a month. Anthropic, Nov 25 2024 and Dec 9 2025. TechCrunch Mar 26 2025 (OpenAI), Apr 9 2025 (Google). Windows Dev Blog May 19 2025. AAIF blog Apr 13 2026 (110M+).
02
With context, days of digging become one sentence .
Who Today With an MCP layer
Operators Toggle 3–5 tools and wait on engineers for answers. Ask in plain language. Get the answer in seconds.
Engineers Hand-build a connector per system, sprint after sprint. One build: a typed, validated MCP server.
CIOs Multi-year data programs before any value. Agent-ready systems in days. Zero changes to them.
Security Fear of code and credentials leaking into AI. Secrets resolved at runtime. Every call audited.
4–6 weeks → 1 day
to make a 3-year-old automotive supply-chain system agent-ready. Zero code changes.
OpenAISep 2026
Anthropic2025–26
MetaSep 2026
xAIAug 2026
Open sourceMIT · 391k+ ★
Open sourceMIT · 251k ★
Python
TypeScript
JavaScript
Go
Java
.NET
Rust
C, C++ and more coming
